Related
I normally program in languages that have the concept of zero-cost abstractions like C++ and Rust.
Currently I'm working in a project that uses C# language. So I was wondering if I can safely create abstractions and higher level code without compromising the performance.
Is that possible in C# or for performance critical code I should just do as low level code as possible?
Just as an example I encountered in my code (don't focus too much on this example, my question is more high level), I needed a function that would return multiple values, for that, my first approach was to use a tuple, so something like this:
public (int, int, float) Function();
or Abstract this tuple into a struct:
public struct Abstraction { int value1; int value2; float value3; };
public Abstraction Function();
What I expected is that the compiler would optimize the Tuple or the Abstraction struct away and simply use the primitive values directly. But what I found is that writing the code using out parameters would improve performance:
public void Function(out int value1, out int value2, out float value3);
I'm guessing the reason is because in the out function, there is no Tuple or Abstraction struct creation.
The problem with the out function version is that I really hate to use parameters as return values, since it seems more like a hack to a language limitation.
So, in the end I'm not sure if I'm just not using the correct configuration so the JIT could use zero-cost abstraction or this is simply not possible or not guaranteed in C#.
First of all, I don't think it makes sense to say that languages "have zero-cost abstractions". Consider the abstraction of function. Is it zero-cost? Generally speaking, it is zero-cost only if it's inlined. And while C++ compilers tend to be really good about inlining functions, they don't inline all functions, so function in C++ is strictly speaking not a zero-cost abstraction. But this difference only matters rarely in practice, which is why you can usually consider a function to be zero-cost.
Now, modern C++ and Rust are designed and implemented in such a way that they make abstractions zero-cost as often as possible. Is this different in C#? Kind of. C# is not designed with such focus on zero-cost abstractions (e.g. invoking a lambda in C# always involves what's effectively a virtual call; invoking a lambda in C++ does not, which makes it much easier to make it zero-cost). Also, JIT compilers generally can't afford to spend as much time on optimizations like inlining and so they generate worse code for abstractions than C++ compilers. (Though this might change in the future, since .Net Core 2.1 introduced a tiered JIT, which means it has more time for optimizations.)
On the other hand, the JIT compiler is tweaked to work well for real code, not for microbenchmarks (which I assume is how you came to the conclusion that returning a struct has worse performance).
In my microbenchmark, using a struct indeed had worse performance, but it was because JIT decided not to inline that version of Function, it was not because of the cost of creating a struct, or anything like that. If I fixed that by using [MethodImpl(MethodImplOptions.AggressiveInlining)], both versions achieved the same performance.
So, returning a struct can be a zero-cost abstraction in C#. Though it's true that there is smaller chance of that happening in C# than in C++.
If you want to know what is the actual effect of switching between out parameters and returning a struct, I suggest you write a more realistic benchmark, not a microbenchmark, and see what the results are. (Assuming I got it right that you used a microbenchmark.)
Yes, you "can"; but is very difficult to control. So, you always had to test and measure.
A practical example with "zero cost abstraction":
using System;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
public class App
{
interface IMessages {
string Welcome{ get; }
string Goodbye { get; }
}
partial struct EnglishMessages : IMessages {
public string Welcome {
get { return "Welcome"; }
}
public string Goodbye {
get { return "Goodbye"; }
}
}
partial struct SpanishMessages : IMessages {
public string Welcome {
get { return "Bienvenido"; }
}
public string Goodbye {
get { return "Adios"; }
}
}
static partial class Messages
{
public static SpanishMessages BuildLang {
get { return default; }
}
}
public static void Main() {
Console.WriteLine(Messages.Welcome);
Console.WriteLine(Messages.Goodbye);
}
static partial class Messages
{
public static string Welcome {
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get { return GetWelcomeFrom(BuildLang); }
}
public static string Goodbye {
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get { return GetGoodbyeFrom(BuildLang); }
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static string GetWelcomeFrom<T>()
where T : struct, IMessages
{
var v = default(T);
return v.Welcome;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static string GetWelcomeFrom<T>(T _)
where T : struct, IMessages
{
return GetWelcomeFrom<T>();
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static string GetGoodbyeFrom<T>()
where T : struct, IMessages
{
var v = default(T);
return v.Goodbye;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static string GetGoodbyeFrom<T>(T _)
where T : struct, IMessages
{
return GetGoodbyeFrom<T>();
}
}
#region
[StructLayout(LayoutKind.Explicit, Size = 0)]
partial struct EnglishMessages { [FieldOffset(0)] int _; }
[StructLayout(LayoutKind.Explicit, Size = 0)]
partial struct SpanishMessages { [FieldOffset(0)] int _; }
#endregion
}
You can able to understand the tricks with this code:
using System;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
public class App
{
interface IMessage {
string Value { get; }
bool IsError { get; }
}
static class Messages
{
// AggressiveInlining increase the inline cost threshold,
// decreased by the use of generics.
//
// This allow inlining because has low cost,
// calculated with the used operations.
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static string GetValue<T>()
where T : struct, IMessage
{
// Problem:
// return default(T).Value
//
// Creates a temporal variable using the CIL stack operations.
// Which avoid some optimizers (like coreclr) to eliminate them.
// Solution:
// Create a variable which is eliminated by the optimizer
// because is unnecessary memory.
var v = default(T);
return v.Value;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static bool IsError<T>()
where T : struct, IMessage
{
var v = default(T);
return v.IsError;
}
}
// The use of partial is only to increase the legibility,
// moving the tricks to the end
partial struct WelcomeMessageEnglish : IMessage {
public string Value {
get { return "Welcome"; }
}
public bool IsError {
get { return false; }
}
}
partial struct WelcomeMessageSpanish : IMessage {
public string Value {
get { return "Bienvenido"; }
}
public bool IsError {
get { return false; }
}
}
public static void Main() {
Console.WriteLine(Messages.GetValue<WelcomeMessageEnglish>() );
Console.WriteLine(Messages.GetValue<WelcomeMessageSpanish>() );
}
// An struct has Size = 1 and is initializated to 0
// This avoid that, setting Size = 0
#region
[StructLayout(LayoutKind.Explicit, Size = 0)]
partial struct WelcomeMessageEnglish { [FieldOffset(0)] int _; }
[StructLayout(LayoutKind.Explicit, Size = 0)]
partial struct WelcomeMessageSpanish { [FieldOffset(0)] int _; }
#endregion
}
I "tested" this in CoreClr, Roslyn, Mono and the abstraction has "zero cost":
App.Main()
L0000: push ebp
L0001: mov ebp, esp
L0003: mov ecx, [0xfd175c4]
L0009: call System.Console.WriteLine(System.String)
L000e: mov ecx, [0xfd17628]
L0014: call System.Console.WriteLine(System.String)
L0019: pop ebp
L001a: ret
For coreclr and roslyn, you can view the asm in SharpLab: Here.
And for mono (in GNU/Linux):
mono --aot zerocost.exe
objdump -d -M intel zerocost.exe.so > zerocost.exe.so.dump
cat zerocost.exe.so.dump #Looking for <App_Main>
When you return something, you always create a new object – you save that step completely when just working "in place" with your out parameters.
Then, you have things that your compiler can't simply optimize away – I'd have to tell you a bit about strict aliasing rules in C, but I don't know C# enough to know whether similar things apply here.
So, in general, creating an object of tuple or Abstraction type is not optimizable away. You specifically specified you wanted to return an object of that type, so that object has to be created by a "general" compilation of the function. You could argue that the compiler knows the context in which Function is called and could infer that it's OK to not generate the object but directly work as if these were reference to the things that you assign the fields of Abstraction to later, but here aliasing rules might get really complicated, and that will in general be logically impossible to do.
I was thrilled to see the new System.Collections.Concurrent namespace in .Net 4.0, quite nice! I've seen ConcurrentDictionary, ConcurrentQueue, ConcurrentStack, ConcurrentBag and BlockingCollection.
One thing that seems to be mysteriously missing is a ConcurrentList<T>. Do I have to write that myself (or get it off the web :) )?
Am I missing something obvious here?
I gave it a try a while back (also: on GitHub). My implementation had some problems, which I won't get into here. Let me tell you, more importantly, what I learned.
Firstly, there's no way you're going to get a full implementation of IList<T> that is lockless and thread-safe. In particular, random insertions and removals are not going to work, unless you also forget about O(1) random access (i.e., unless you "cheat" and just use some sort of linked list and let the indexing suck).
What I thought might be worthwhile was a thread-safe, limited subset of IList<T>: in particular, one that would allow an Add and provide random read-only access by index (but no Insert, RemoveAt, etc., and also no random write access).
This was the goal of my ConcurrentList<T> implementation. But when I tested its performance in multithreaded scenarios, I found that simply synchronizing adds to a List<T> was faster. Basically, adding to a List<T> is lightning fast already; the complexity of the computational steps involved is miniscule (increment an index and assign to an element in an array; that's really it). You would need a ton of concurrent writes to see any sort of lock contention on this; and even then, the average performance of each write would still beat out the more expensive albeit lockless implementation in ConcurrentList<T>.
In the relatively rare event that the list's internal array needs to resize itself, you do pay a small cost. So ultimately I concluded that this was the one niche scenario where an add-only ConcurrentList<T> collection type would make sense: when you want guaranteed low overhead of adding an element on every single call (so, as opposed to an amortized performance goal).
It's simply not nearly as useful a class as you would think.
What would you use a ConcurrentList for?
The concept of a Random Access container in a threaded world isn't as useful as it may appear. The statement
if (i < MyConcurrentList.Count)
x = MyConcurrentList[i];
as a whole would still not be thread-safe.
Instead of creating a ConcurrentList, try to build solutions with what's there. The most common classes are the ConcurrentBag and especially the BlockingCollection.
With all due respect to the great answers provided already, there are times that I simply want a thread-safe IList. Nothing advanced or fancy. Performance is important in many cases but at times that just isn't a concern. Yes, there are always going to be challenges without methods like "TryGetValue" etc, but most cases I just want something that I can enumerate without needing to worry about putting locks around everything. And yes, somebody can probably find some "bug" in my implementation that might lead to a deadlock or something (I suppose) but lets be honest: When it comes to multi-threading, if you don't write your code correctly, it is going deadlock anyway. With that in mind I decided to make a simple ConcurrentList implementation that provides these basic needs.
And for what its worth: I did a basic test of adding 10,000,000 items to regular List and ConcurrentList and the results were:
List finished in: 7793 milliseconds.
Concurrent finished in: 8064 milliseconds.
public class ConcurrentList<T> : IList<T>, IDisposable
{
#region Fields
private readonly List<T> _list;
private readonly ReaderWriterLockSlim _lock;
#endregion
#region Constructors
public ConcurrentList()
{
this._lock = new ReaderWriterLockSlim(LockRecursionPolicy.NoRecursion);
this._list = new List<T>();
}
public ConcurrentList(int capacity)
{
this._lock = new ReaderWriterLockSlim(LockRecursionPolicy.NoRecursion);
this._list = new List<T>(capacity);
}
public ConcurrentList(IEnumerable<T> items)
{
this._lock = new ReaderWriterLockSlim(LockRecursionPolicy.NoRecursion);
this._list = new List<T>(items);
}
#endregion
#region Methods
public void Add(T item)
{
try
{
this._lock.EnterWriteLock();
this._list.Add(item);
}
finally
{
this._lock.ExitWriteLock();
}
}
public void Insert(int index, T item)
{
try
{
this._lock.EnterWriteLock();
this._list.Insert(index, item);
}
finally
{
this._lock.ExitWriteLock();
}
}
public bool Remove(T item)
{
try
{
this._lock.EnterWriteLock();
return this._list.Remove(item);
}
finally
{
this._lock.ExitWriteLock();
}
}
public void RemoveAt(int index)
{
try
{
this._lock.EnterWriteLock();
this._list.RemoveAt(index);
}
finally
{
this._lock.ExitWriteLock();
}
}
public int IndexOf(T item)
{
try
{
this._lock.EnterReadLock();
return this._list.IndexOf(item);
}
finally
{
this._lock.ExitReadLock();
}
}
public void Clear()
{
try
{
this._lock.EnterWriteLock();
this._list.Clear();
}
finally
{
this._lock.ExitWriteLock();
}
}
public bool Contains(T item)
{
try
{
this._lock.EnterReadLock();
return this._list.Contains(item);
}
finally
{
this._lock.ExitReadLock();
}
}
public void CopyTo(T[] array, int arrayIndex)
{
try
{
this._lock.EnterReadLock();
this._list.CopyTo(array, arrayIndex);
}
finally
{
this._lock.ExitReadLock();
}
}
public IEnumerator<T> GetEnumerator()
{
return new ConcurrentEnumerator<T>(this._list, this._lock);
}
IEnumerator IEnumerable.GetEnumerator()
{
return new ConcurrentEnumerator<T>(this._list, this._lock);
}
~ConcurrentList()
{
this.Dispose(false);
}
public void Dispose()
{
this.Dispose(true);
}
private void Dispose(bool disposing)
{
if (disposing)
GC.SuppressFinalize(this);
this._lock.Dispose();
}
#endregion
#region Properties
public T this[int index]
{
get
{
try
{
this._lock.EnterReadLock();
return this._list[index];
}
finally
{
this._lock.ExitReadLock();
}
}
set
{
try
{
this._lock.EnterWriteLock();
this._list[index] = value;
}
finally
{
this._lock.ExitWriteLock();
}
}
}
public int Count
{
get
{
try
{
this._lock.EnterReadLock();
return this._list.Count;
}
finally
{
this._lock.ExitReadLock();
}
}
}
public bool IsReadOnly
{
get { return false; }
}
#endregion
}
public class ConcurrentEnumerator<T> : IEnumerator<T>
{
#region Fields
private readonly IEnumerator<T> _inner;
private readonly ReaderWriterLockSlim _lock;
#endregion
#region Constructor
public ConcurrentEnumerator(IEnumerable<T> inner, ReaderWriterLockSlim #lock)
{
this._lock = #lock;
this._lock.EnterReadLock();
this._inner = inner.GetEnumerator();
}
#endregion
#region Methods
public bool MoveNext()
{
return _inner.MoveNext();
}
public void Reset()
{
_inner.Reset();
}
public void Dispose()
{
this._lock.ExitReadLock();
}
#endregion
#region Properties
public T Current
{
get { return _inner.Current; }
}
object IEnumerator.Current
{
get { return _inner.Current; }
}
#endregion
}
The reason why there is no ConcurrentList is because it fundamentally cannot be written. The reason why is that several important operations in IList rely on indices, and that just plain won't work. For example:
int catIndex = list.IndexOf("cat");
list.Insert(catIndex, "dog");
The effect that the author is going after is to insert "dog" before "cat", but in a multithreaded environment, anything can happen to the list between those two lines of code. For example, another thread might do list.RemoveAt(0), shifting the entire list to the left, but crucially, catIndex will not change. The impact here is that the Insert operation will actually put the "dog" after the cat, not before it.
The several implementations that you see offered as "answers" to this question are well-meaning, but as the above shows, they don't offer reliable results. If you really want list-like semantics in a multithreaded environment, you can't get there by putting locks inside the list implementation methods. You have to ensure that any index you use lives entirely inside the context of the lock. The upshot is that you can use a List in a multithreaded environment with the right locking, but the list itself cannot be made to exist in that world.
If you think you need a concurrent list, there are really just two possibilities:
What you really need is a ConcurrentBag
You need to create your own collection, perhaps implemented with a List and your own concurrency control.
If you have a ConcurrentBag and are in a position where you need to pass it as an IList, then you have a problem, because the method you're calling has specified that they might try to do something like I did above with the cat & dog. In most worlds, what that means is that the method you're calling is simply not built to work in a multi-threaded environment. That means you either refactor it so that it is or, if you can't, you're going to have to handle it very carefully. You you'll almost certainly be required to create your own collection with its own locks, and call the offending method within a lock.
ConcurrentList (as a resizeable array, not a linked list) is not easy to write with nonblocking operations. Its API doesn't translate well to a "concurrent" version.
In cases where reads greatly outnumber writes, or (however frequent) writes are non-concurrent, a copy-on-write approach may be appropriate.
The implementation shown below is
lockless
blazingly fast for concurrent reads, even while concurrent modifications are ongoing - no matter how long they take
because "snapshots" are immutable, lockless atomicity is possible, i.e. var snap = _list; snap[snap.Count - 1]; will never (well, except for an empty list of course) throw, and you also get thread-safe enumeration with snapshot semantics for free.. how I LOVE immutability!
implemented generically, applicable to any data structure and any type of modification
dead simple, i.e. easy to test, debug, verify by reading the code
usable in .Net 3.5
For copy-on-write to work, you have to keep your data structures effectively immutable, i.e. no one is allowed to change them after you made them available to other threads. When you want to modify, you
clone the structure
make modifications on the clone
atomically swap in the reference to the modified clone
Code
static class CopyOnWriteSwapper
{
public static void Swap<T>(ref T obj, Func<T, T> cloner, Action<T> op)
where T : class
{
while (true)
{
var objBefore = Volatile.Read(ref obj);
var newObj = cloner(objBefore);
op(newObj);
if (Interlocked.CompareExchange(ref obj, newObj, objBefore) == objBefore)
return;
}
}
}
Usage
CopyOnWriteSwapper.Swap(ref _myList,
orig => new List<string>(orig),
clone => clone.Add("asdf"));
If you need more performance, it will help to ungenerify the method, e.g. create one method for every type of modification (Add, Remove, ...) you want, and hard code the function pointers cloner and op.
N.B. #1 It is your responsibility to make sure the no one modifies the (supposedly) immutable data structure. There's nothing we can do in a generic implementation to prevent that, but when specializing to List<T>, you could guard against modification using List.AsReadOnly()
N.B. #2 Be careful about the values in the list. The copy on write approach above guards their list membership only, but if you'd put not strings, but some other mutable objects in there, you have to take care of thread safety (e.g. locking). But that is orthogonal to this solution and e.g. locking of the mutable values can be easily used without issues. You just need to be aware of it.
N.B. #3 If your data structure is huge and you modify it frequently, the copy-all-on-write approach might be prohibitive both in terms of memory consumption and the CPU cost of copying involved. In that case, you might want to use MS's Immutable Collections instead.
System.Collections.Generic.List<t> is already thread safe for multiple readers. Trying to make it thread safe for multiple writers wouldn't make sense. (For reasons Henk and Stephen already mentioned)
Some people hilighted some goods points (and some of my thoughts):
It could looklikes insane to unable random accesser (indexer) but to me it appears fine. You only have to think that there is many methods on multi-threaded collections that could fail like Indexer and Delete. You could also define failure (fallback) action for write accessor like "fail" or simply "add at the end".
It is not because it is a multithreaded collection that it will always be used in a multithreaded context. Or it could also be used by only one writer and one reader.
Another way to be able to use indexer in a safe manner could be to wrap actions into a lock of the collection using its root (if made public).
For many people, making a rootLock visible goes agaist "Good practice". I'm not 100% sure about this point because if it is hidden you remove a lot of flexibility to the user. We always have to remember that programming multithread is not for anybody. We can't prevent every kind of wrong usage.
Microsoft will have to do some work and define some new standard to introduce proper usage of Multithreaded collection. First the IEnumerator should not have a moveNext but should have a GetNext that return true or false and get an out paramter of type T (this way the iteration would not be blocking anymore). Also, Microsoft already use "using" internally in the foreach but sometimes use the IEnumerator directly without wrapping it with "using" (a bug in collection view and probably at more places) - Wrapping usage of IEnumerator is a recommended pratice by Microsoft. This bug remove good potential for safe iterator... Iterator that lock collection in constructor and unlock on its Dispose method - for a blocking foreach method.
That is not an answer. This is only comments that do not really fit to a specific place.
... My conclusion, Microsoft has to make some deep changes to the "foreach" to make MultiThreaded collection easier to use. Also it has to follow there own rules of IEnumerator usage. Until that, we can write a MultiThreadList easily that would use a blocking iterator but that will not follow "IList". Instead, you will have to define own "IListPersonnal" interface that could fail on "insert", "remove" and random accessor (indexer) without exception. But who will want to use it if it is not standard ?
I implemented one similar to Brian's. Mine is different:
I manage the array directly.
I don't enter the locks within the try block.
I use yield return for producing an enumerator.
I support lock recursion. This allows reads from list during iteration.
I use upgradable read locks where possible.
DoSync and GetSync methods allowing sequential interactions that require exclusive access to the list.
The code:
public class ConcurrentList<T> : IList<T>, IDisposable
{
private ReaderWriterLockSlim _lock = new ReaderWriterLockSlim(LockRecursionPolicy.SupportsRecursion);
private int _count = 0;
public int Count
{
get
{
_lock.EnterReadLock();
try
{
return _count;
}
finally
{
_lock.ExitReadLock();
}
}
}
public int InternalArrayLength
{
get
{
_lock.EnterReadLock();
try
{
return _arr.Length;
}
finally
{
_lock.ExitReadLock();
}
}
}
private T[] _arr;
public ConcurrentList(int initialCapacity)
{
_arr = new T[initialCapacity];
}
public ConcurrentList():this(4)
{ }
public ConcurrentList(IEnumerable<T> items)
{
_arr = items.ToArray();
_count = _arr.Length;
}
public void Add(T item)
{
_lock.EnterWriteLock();
try
{
var newCount = _count + 1;
EnsureCapacity(newCount);
_arr[_count] = item;
_count = newCount;
}
finally
{
_lock.ExitWriteLock();
}
}
public void AddRange(IEnumerable<T> items)
{
if (items == null)
throw new ArgumentNullException("items");
_lock.EnterWriteLock();
try
{
var arr = items as T[] ?? items.ToArray();
var newCount = _count + arr.Length;
EnsureCapacity(newCount);
Array.Copy(arr, 0, _arr, _count, arr.Length);
_count = newCount;
}
finally
{
_lock.ExitWriteLock();
}
}
private void EnsureCapacity(int capacity)
{
if (_arr.Length >= capacity)
return;
int doubled;
checked
{
try
{
doubled = _arr.Length * 2;
}
catch (OverflowException)
{
doubled = int.MaxValue;
}
}
var newLength = Math.Max(doubled, capacity);
Array.Resize(ref _arr, newLength);
}
public bool Remove(T item)
{
_lock.EnterUpgradeableReadLock();
try
{
var i = IndexOfInternal(item);
if (i == -1)
return false;
_lock.EnterWriteLock();
try
{
RemoveAtInternal(i);
return true;
}
finally
{
_lock.ExitWriteLock();
}
}
finally
{
_lock.ExitUpgradeableReadLock();
}
}
public IEnumerator<T> GetEnumerator()
{
_lock.EnterReadLock();
try
{
for (int i = 0; i < _count; i++)
// deadlocking potential mitigated by lock recursion enforcement
yield return _arr[i];
}
finally
{
_lock.ExitReadLock();
}
}
IEnumerator IEnumerable.GetEnumerator()
{
return this.GetEnumerator();
}
public int IndexOf(T item)
{
_lock.EnterReadLock();
try
{
return IndexOfInternal(item);
}
finally
{
_lock.ExitReadLock();
}
}
private int IndexOfInternal(T item)
{
return Array.FindIndex(_arr, 0, _count, x => x.Equals(item));
}
public void Insert(int index, T item)
{
_lock.EnterUpgradeableReadLock();
try
{
if (index > _count)
throw new ArgumentOutOfRangeException("index");
_lock.EnterWriteLock();
try
{
var newCount = _count + 1;
EnsureCapacity(newCount);
// shift everything right by one, starting at index
Array.Copy(_arr, index, _arr, index + 1, _count - index);
// insert
_arr[index] = item;
_count = newCount;
}
finally
{
_lock.ExitWriteLock();
}
}
finally
{
_lock.ExitUpgradeableReadLock();
}
}
public void RemoveAt(int index)
{
_lock.EnterUpgradeableReadLock();
try
{
if (index >= _count)
throw new ArgumentOutOfRangeException("index");
_lock.EnterWriteLock();
try
{
RemoveAtInternal(index);
}
finally
{
_lock.ExitWriteLock();
}
}
finally
{
_lock.ExitUpgradeableReadLock();
}
}
private void RemoveAtInternal(int index)
{
Array.Copy(_arr, index + 1, _arr, index, _count - index-1);
_count--;
// release last element
Array.Clear(_arr, _count, 1);
}
public void Clear()
{
_lock.EnterWriteLock();
try
{
Array.Clear(_arr, 0, _count);
_count = 0;
}
finally
{
_lock.ExitWriteLock();
}
}
public bool Contains(T item)
{
_lock.EnterReadLock();
try
{
return IndexOfInternal(item) != -1;
}
finally
{
_lock.ExitReadLock();
}
}
public void CopyTo(T[] array, int arrayIndex)
{
_lock.EnterReadLock();
try
{
if(_count > array.Length - arrayIndex)
throw new ArgumentException("Destination array was not long enough.");
Array.Copy(_arr, 0, array, arrayIndex, _count);
}
finally
{
_lock.ExitReadLock();
}
}
public bool IsReadOnly
{
get { return false; }
}
public T this[int index]
{
get
{
_lock.EnterReadLock();
try
{
if (index >= _count)
throw new ArgumentOutOfRangeException("index");
return _arr[index];
}
finally
{
_lock.ExitReadLock();
}
}
set
{
_lock.EnterUpgradeableReadLock();
try
{
if (index >= _count)
throw new ArgumentOutOfRangeException("index");
_lock.EnterWriteLock();
try
{
_arr[index] = value;
}
finally
{
_lock.ExitWriteLock();
}
}
finally
{
_lock.ExitUpgradeableReadLock();
}
}
}
public void DoSync(Action<ConcurrentList<T>> action)
{
GetSync(l =>
{
action(l);
return 0;
});
}
public TResult GetSync<TResult>(Func<ConcurrentList<T>,TResult> func)
{
_lock.EnterWriteLock();
try
{
return func(this);
}
finally
{
_lock.ExitWriteLock();
}
}
public void Dispose()
{
_lock.Dispose();
}
}
In sequentially executing code the data structures used are different from (well written) concurrently executing code. The reason is that sequential code implies implicit order. Concurrent code however does not imply any order; better yet it implies the lack of any defined order!
Due to this, data structures with implied order (like List) are not very useful for solving concurrent problems. A list implies order, but it does not clearly define what that order is. Because of this the execution order of the code manipulating the list will determine (to some degree) the implicit order of the list, which is in direct conflict with an efficient concurrent solution.
Remember concurrency is a data problem, not a code problem! You cannot Implement the code first (or rewriting existing sequential code) and get a well designed concurrent solution. You need to design the data structures first while keeping in mind that implicit ordering doesn’t exist in a concurrent system.
lockless Copy and Write approach works great if you're not dealing with too many items.
Here's a class I wrote:
public class CopyAndWriteList<T>
{
public static List<T> Clear(List<T> list)
{
var a = new List<T>(list);
a.Clear();
return a;
}
public static List<T> Add(List<T> list, T item)
{
var a = new List<T>(list);
a.Add(item);
return a;
}
public static List<T> RemoveAt(List<T> list, int index)
{
var a = new List<T>(list);
a.RemoveAt(index);
return a;
}
public static List<T> Remove(List<T> list, T item)
{
var a = new List<T>(list);
a.Remove(item);
return a;
}
}
example usage:
orders_BUY = CopyAndWriteList.Clear(orders_BUY);
I'm surprised no-one has mentioned using LinkedList as a base for writing a specialised class.
Often we don't need the full API's of the various collection classes, and if you write mostly functional side effect free code, using immutable classes as far as possible, then you'll actually NOT want to mutate the collection favouring various snapshot implementations.
LinkedList solves some difficult problems of creating snapshot copies/clones of large collections. I also use it to create "threadsafe" enumerators to enumerate over the collection. I can cheat, because I know that I'm not changing the collection in any way other than appending, I can keep track of the list size, and only lock on changes to list size. Then my enumerator code simply enumerates from 0 to n for any thread that wants a "snapshot" of the append only collection, that will be guaranteed to represent a "snapshot" of the collection at any moment in time, regardless of what other threads are appending to the head of the collection.
I'm pretty certain that most requirements are often extremely simple, and you need 2 or 3 methods only. Writing a truly generic library is awfully difficult, but solving your own codes needs can sometimes be easy with a trick or two.
Long live LinkedList and good functional programming.
Cheers, ... love ya all!
Al
p.s. sample hack AppendOnly class here : https://github.com/goblinfactory/AppendOnly
I want an implementation of List<T> as a property which can be used thread-safely without any doubt.
Something like this:
private List<T> _list;
private List<T> MyT
{
get { // return a copy of _list; }
set { _list = value; }
}
It seems still I need to return a copy (cloned) of collection so if somewhere we are iterating the collection and at the same time the collection is set, then no exception is raised.
How to implement a thread-safe collection property?
If you are targetting .Net 4 there are a few options in System.Collections.Concurrent Namespace
You could use ConcurrentBag<T> in this case instead of List<T>
Even as it got the most votes, one usually can't take System.Collections.Concurrent.ConcurrentBag<T> as a thread-safe replacement for System.Collections.Generic.List<T> as it is (Radek Stromský already pointed it out) not ordered.
But there is a class called System.Collections.Generic.SynchronizedCollection<T> that is already since .NET 3.0 part of the framework, but it is that well hidden in a location where one does not expect it that it is little known and probably you have never ever stumbled over it (at least I never did).
SynchronizedCollection<T> is compiled into assembly System.ServiceModel.dll (which is part of the client profile but not of the portable class library).
I would think making a sample ThreadSafeList class would be easy:
public class ThreadSafeList<T> : IList<T>
{
protected List<T> _internalList = new List<T>();
// Other Elements of IList implementation
public IEnumerator<T> GetEnumerator()
{
return Clone().GetEnumerator();
}
System.Collections.IEnumerator System.Collections.IEnumerable.GetEnumerator()
{
return Clone().GetEnumerator();
}
protected static object _lock = new object();
public List<T> Clone()
{
List<T> newList = new List<T>();
lock (_lock)
{
_internalList.ForEach(x => newList.Add(x));
}
return newList;
}
}
You simply clone the list before requesting an enumerator, and thus any enumeration is working off a copy that can't be modified while running.
Even accepted answer is ConcurrentBag, I don't think it's real replacement of list in all cases, as Radek's comment to the answer says: "ConcurrentBag is unordered collection, so unlike List it does not guarantee ordering. Also you cannot access items by index".
So if you use .NET 4.0 or higher, a workaround could be to use ConcurrentDictionary with integer TKey as array index and TValue as array value. This is recommended way of replacing list in Pluralsight's C# Concurrent Collections course. ConcurrentDictionary solves both problems mentioned above: index accessing and ordering (we can not rely on ordering as it's hash table under the hood, but current .NET implementation saves order of elements adding).
C#'s ArrayList class has a Synchronized method.
var threadSafeArrayList = ArrayList.Synchronized(new ArrayList());
This returns a thread safe wrapper around any instance of IList. All operations need to be performed through the wrapper to ensure thread safety.
In .NET Core (any version), you can use ImmutableList, which has all the functionality of List<T>.
If you look at the source code for List of T (https://referencesource.microsoft.com/#mscorlib/system/collections/generic/list.cs,c66df6f36c131877) you will notice there is a class there (which is of course internal - why, Microsoft, why?!?!) called SynchronizedList of T. I am copy pasting the code here:
[Serializable()]
internal class SynchronizedList : IList<T> {
private List<T> _list;
private Object _root;
internal SynchronizedList(List<T> list) {
_list = list;
_root = ((System.Collections.ICollection)list).SyncRoot;
}
public int Count {
get {
lock (_root) {
return _list.Count;
}
}
}
public bool IsReadOnly {
get {
return ((ICollection<T>)_list).IsReadOnly;
}
}
public void Add(T item) {
lock (_root) {
_list.Add(item);
}
}
public void Clear() {
lock (_root) {
_list.Clear();
}
}
public bool Contains(T item) {
lock (_root) {
return _list.Contains(item);
}
}
public void CopyTo(T[] array, int arrayIndex) {
lock (_root) {
_list.CopyTo(array, arrayIndex);
}
}
public bool Remove(T item) {
lock (_root) {
return _list.Remove(item);
}
}
System.Collections.IEnumerator System.Collections.IEnumerable.GetEnumerator() {
lock (_root) {
return _list.GetEnumerator();
}
}
IEnumerator<T> IEnumerable<T>.GetEnumerator() {
lock (_root) {
return ((IEnumerable<T>)_list).GetEnumerator();
}
}
public T this[int index] {
get {
lock(_root) {
return _list[index];
}
}
set {
lock(_root) {
_list[index] = value;
}
}
}
public int IndexOf(T item) {
lock (_root) {
return _list.IndexOf(item);
}
}
public void Insert(int index, T item) {
lock (_root) {
_list.Insert(index, item);
}
}
public void RemoveAt(int index) {
lock (_root) {
_list.RemoveAt(index);
}
}
}
Personally I think they knew a better implementation using SemaphoreSlim could be created, but didn't get to it.
I would suggest anyone dealing with a List<T> in multi-threading scenarios to take look at Immutable Collections in particular the ImmutableArray.
I've found it very useful when you have:
Relatively few items in the list
Not so many read/write operations
A LOT of concurrent access (i.e. many threads that access the list in reading mode)
Also can be useful when you need to implement some sort of transaction-like behavior (i.e. revert an insert/update/delete operation in case of fail)
It seems like many of the people finding this are wanting a thread safe indexed dynamically sized collection. The closest and easiest thing I know of would be.
System.Collections.Concurrent.ConcurrentDictionary<int, YourDataType>
This would require you to ensure your key is properly incremented if you want normal indexing behavior. If you are careful .count() could suffice as the key for any new key value pairs you add.
You can also use the more primitive
Monitor.Enter(lock);
Monitor.Exit(lock);
which lock uses (see this post C# Locking an object that is reassigned in lock block).
If you are expecting exceptions in the code this is not safe but it allows you to do something like the following:
using System;
using System.Collections.Generic;
using System.Threading;
using System.Linq;
public class Something
{
private readonly object _lock;
private readonly List<string> _contents;
public Something()
{
_lock = new object();
_contents = new List<string>();
}
public Modifier StartModifying()
{
return new Modifier(this);
}
public class Modifier : IDisposable
{
private readonly Something _thing;
public Modifier(Something thing)
{
_thing = thing;
Monitor.Enter(Lock);
}
public void OneOfLotsOfDifferentOperations(string input)
{
DoSomethingWith(input);
}
private void DoSomethingWith(string input)
{
Contents.Add(input);
}
private List<string> Contents
{
get { return _thing._contents; }
}
private object Lock
{
get { return _thing._lock; }
}
public void Dispose()
{
Monitor.Exit(Lock);
}
}
}
public class Caller
{
public void Use(Something thing)
{
using (var modifier = thing.StartModifying())
{
modifier.OneOfLotsOfDifferentOperations("A");
modifier.OneOfLotsOfDifferentOperations("B");
modifier.OneOfLotsOfDifferentOperations("A");
modifier.OneOfLotsOfDifferentOperations("A");
modifier.OneOfLotsOfDifferentOperations("A");
}
}
}
One of the nice things about this is you'll get the lock for the duration of the series of operations (rather than locking in each operation). Which means that the output should come out in the right chunks (my usage of this was getting some output onto screen from an external process)
I do really like the simplicity + transparency of the ThreadSafeList + that does the important bit in stopping crashes
I believe _list.ToList() will make you a copy. You can also query it if you need to such as :
_list.Select("query here").ToList();
Anyways, msdn says this is indeed a copy and not simply a reference. Oh, and yes, you will need to lock in the set method as the others have pointed out.
Looking at the original sample one may guess that the intention was to be able to simply replace the list with the new one. The setter on the property tells us about it.
The Micrisoft's Thread-Safe Collections are for safely adding and removing items from collection. But if in the application logic you are intending to replace the collection with the new one, one may guess, again, that the adding and deleting functionality of the List is not required.
If this is the case then, the simple answer would be to use IReadOnlyList interface:
private IReadOnlyList<T> _readOnlyList = new List<T>();
private IReadOnlyList<T> MyT
{
get { return _readOnlyList; }
set { _readOnlyList = value; }
}
One doesn't need to use any locking in this situation because there is no way to modify the collection. If in the setter the "_readOnlyList = value;" will be replaced by something more complicated then the lock could be required.
Basically if you want to enumerate safely, you need to use lock.
Please refer to MSDN on this. http://msdn.microsoft.com/en-us/library/6sh2ey19.aspx
Here is part of MSDN that you might be interested:
Public static (Shared in Visual Basic) members of this type are thread safe. Any instance members are not guaranteed to be thread safe.
A List can support multiple readers concurrently, as long as the collection is not modified. Enumerating through a collection is intrinsically not a thread-safe procedure. In the rare case where an enumeration contends with one or more write accesses, the only way to ensure thread safety is to lock the collection during the entire enumeration. To allow the collection to be accessed by multiple threads for reading and writing, you must implement your own synchronization.
Here is the class for thread safe list without lock
public class ConcurrentList
{
private long _i = 1;
private ConcurrentDictionary<long, T> dict = new ConcurrentDictionary<long, T>();
public int Count()
{
return dict.Count;
}
public List<T> ToList()
{
return dict.Values.ToList();
}
public T this[int i]
{
get
{
long ii = dict.Keys.ToArray()[i];
return dict[ii];
}
}
public void Remove(T item)
{
T ov;
var dicItem = dict.Where(c => c.Value.Equals(item)).FirstOrDefault();
if (dicItem.Key > 0)
{
dict.TryRemove(dicItem.Key, out ov);
}
this.CheckReset();
}
public void RemoveAt(int i)
{
long v = dict.Keys.ToArray()[i];
T ov;
dict.TryRemove(v, out ov);
this.CheckReset();
}
public void Add(T item)
{
dict.TryAdd(_i, item);
_i++;
}
public IEnumerable<T> Where(Func<T, bool> p)
{
return dict.Values.Where(p);
}
public T FirstOrDefault(Func<T, bool> p)
{
return dict.Values.Where(p).FirstOrDefault();
}
public bool Any(Func<T, bool> p)
{
return dict.Values.Where(p).Count() > 0 ? true : false;
}
public void Clear()
{
dict.Clear();
}
private void CheckReset()
{
if (dict.Count == 0)
{
this.Reset();
}
}
private void Reset()
{
_i = 1;
}
}
Use the lock statement to do this. (Read here for more information.)
private List<T> _list;
private List<T> MyT
{
get { return _list; }
set
{
//Lock so only one thread can change the value at any given time.
lock (_list)
{
_list = value;
}
}
}
FYI this probably isn't exactly what your asking - you likely want to lock farther out in your code but I can't assume that. Have a look at the lock keyword and tailor its use to your specific situation.
If you need to, you could lock in both the get and set block using the _list variable which would make it so a read/write can not occur at the same time.
I have a simple method that uses an iterator block to return an IEnumerable<T>:
IEnumerable<MyItem> GetItems()
{
foreach (var item in Items)
{
yield return item;
}
}
Ordinarily, this method works fine, but if I apply a [SecurityCritical] attribute to the assembly (or to the class that contains the above method), it throws a TypeLoadException when attempting to invoke the method. The type that is failing to load is the compiler-generated class that corresponds to the iterator method, and it is its GetEnumerator method that is causing the problem, since it is security transparent.
For comparison, if I modify the above method so that it populates and returns a List<MyItem>, everything works fine.
Any suggestions?
Thanks,
Tim.
It isn't the neatest thing to do, so hopefully you can find a better way, but you could always forgo the compiler-generated code and create your own class that implements IEnumerator<MyItem> (and perhaps your own class implementing IEnumerable<MyItem> - depending on complexity, doing so may make things easier or more difficult), and then build the enumerator more or less as you would in the days before .NET2.0.
If the logic of your real iterator block is very complicated, you might find looking at the reflection of the class the compiler created for you to be a good starting point in doing this, though sometimes the generated code is more complicated (or at least, less readable) than the approach one would take oneself.
It's always a bit disappointing to have to build an IEnumerator class when yield has made it so nice for us 99% of the time, but there are still times when its necessary, and it might solve your problem here.
I had the very same problem, in a complicated application. Spring comes in between and said that the 'blahblah' type is not Serializable and sure it was correct, Here is the disassembled code of compiler generated code and sure it's not Serializable. Maybe this was your problem too, and the solution is what you mentioned yourself cause the List is actually a Serializable type.
The code generate for yield return new KeyValuePair<??? ???>(???,???);
[CompilerGenerated, DebuggerDisplay(#"\{ x = {x}, y = {y} }", Type="<Anonymous Type>")]
internal sealed class <>f__AnonymousType0<<x>j__TPar, <y>j__TPar>
{
// Fields
[DebuggerBrowsable(DebuggerBrowsableState.Never)]
private readonly <x>j__TPar <x>i__Field;
[DebuggerBrowsable(DebuggerBrowsableState.Never)]
private readonly <y>j__TPar <y>i__Field;
// Methods
[DebuggerHidden]
public <>f__AnonymousType0(<x>j__TPar x, <y>j__TPar y)
{
this.<x>i__Field = x;
this.<y>i__Field = y;
}
[DebuggerHidden]
public override bool Equals(object value)
{
var type = value as <>f__AnonymousType0<<x>j__TPar, <y>j__TPar>;
return (((type != null) && EqualityComparer<<x>j__TPar>.Default.Equals(this.<x>i__Field, type.<x>i__Field)) && EqualityComparer<<y>j__TPar>.Default.Equals(this.<y>i__Field, type.<y>i__Field));
}
[DebuggerHidden]
public override int GetHashCode()
{
int num = -576933007;
num = (-1521134295 * num) + EqualityComparer<<x>j__TPar>.Default.GetHashCode(this.<x>i__Field);
return ((-1521134295 * num) + EqualityComparer<<y>j__TPar>.Default.GetHashCode(this.<y>i__Field));
}
[DebuggerHidden]
public override string ToString()
{
StringBuilder builder = new StringBuilder();
builder.Append("{ x = ");
builder.Append(this.<x>i__Field);
builder.Append(", y = ");
builder.Append(this.<y>i__Field);
builder.Append(" }");
return builder.ToString();
}
// Properties
public <x>j__TPar x
{
get
{
return this.<x>i__Field;
}
}
public <y>j__TPar y
{
get
{
return this.<y>i__Field;
}
}
}
You can vote for this issue: https://connect.microsoft.com/VisualStudio/feedback/details/667328/yield-and-securitycriticalattribute-problem
[EDIT] Response from Microsoft:
We've looked at SecurityCritical iterators and decided not to try to
make that work at least for this release. It is a significant and
complicated effort, and it does not seem too useful, as the call
through IEnumerator.MoveNext would be calling through a non-critical
interface.
We'll probably revisit this again in a later release; especially if we
see common scenarios for it.
I am porting a library from C++ to C#. The old library uses vectors from C++ and in the C# I am using generic Dictionaries because they're actually a good data structure for what I'm doing (each element has an ID, then I just use using TypeDictionary = Dictionary<String, Type>;). Now, in the C# code I use a loop like this one
TypeDictionary.Enumerator tdEnum = MyTypeDictionary.GetEnumerator();
while( tdEnum.MoveNext() )
{
Type element = typeElement.Current.Value;
// More code here
}
to iterate through the elements of the collection. The problem is that in particular cases I need to check if a certain enumerator has reached the end of the collection, in C++ I would have done a check like this:
if ( tdEnum == MyTypeDictionary.end() ) // More code here
But I just don't know how to handle this situation in C#, any ideas?
Thank you
Tommaso
Here's a pretty simple way of accomplishing this.
bool hasNext = tdEnum.MoveNext();
while (hasNext) {
int i = tdEnum.Current;
hasNext = tdEnum.MoveNext();
}
I found an online tutorial that also may help you understand how this works.
http://www.c-sharpcorner.com/UploadFile/prasadh/Enumerators11132005232321PM/Enumerators.aspx
You know that you're at the end of an iterator when MoveNext() returns false. Otherwise you need to upgrade to a more descriptive data structure like IList<T>.
I have a "smart iterator" class in MiscUtil which you may find useful. It lets you test whether you're currently looking at the start or end of the sequence, and the index within the sequence. See the usage page for more information.
Of course in most cases you can just get away with doing this manually using the result of MoveNext(), but occasionally the extra encapsulation comes in handy.
Note that by necessity, this iterator will always have actually consumed one more value than it's yielded, in order to know whether or not it's reached the end. In most cases that isn't an issue, but it could occasionally give some odd experiences when debugging.
Using the decorator pattern to hold a value if the enumerator has ended is a valid approach.
Since it implements IEnumerator, you won't find difficulties to replace it in your code.
Here's a test class:
using System.Collections.Generic;
using Microsoft.VisualStudio.TestTools.UnitTesting;
using MyDictionary = System.Collections.Generic.Dictionary<int, string>;
using MyKeyValue = System.Collections.Generic.KeyValuePair<int, string>;
namespace TestEnumerator
{
[TestClass]
public class UnitTest1
{
[TestMethod]
public void TestingMyEnumeradorPlus()
{
var itens = new MyDictionary()
{
{ 1, "aaa" },
{ 2, "bbb" }
};
var enumerator = new EnumeradorPlus<MyKeyValue>(itens.GetEnumerator());
enumerator.MoveNext();
Assert.IsFalse(enumerator.Ended);
enumerator.MoveNext();
Assert.IsFalse(enumerator.Ended);
enumerator.MoveNext();
Assert.IsTrue(enumerator.Ended);
}
}
public class EnumeradorPlus<T> : IEnumerator<T>
{
private IEnumerator<T> _internal;
private bool _hasEnded = false;
public EnumeradorPlus(IEnumerator<T> enumerator)
{
_internal = enumerator;
}
public T Current
{
get { return _internal.Current; }
}
public void Dispose()
{
_internal.Dispose();
}
object System.Collections.IEnumerator.Current
{
get { return _internal.Current; }
}
public bool MoveNext()
{
bool moved = _internal.MoveNext();
if (!moved)
_hasEnded = true;
return moved;
}
public void Reset()
{
_internal.Reset();
_hasEnded = false;
}
public bool Ended
{
get { return _hasEnded; }
}
}
}
Coming from C++ you might not be up to date on C# syntax. Perhaps you could simply use the foreach construct to avoid the test all together. The following code will be executed once for each element in your dictionary:
foreach (var element in MyTypeDictionary)
{
// More code here
}