Delphi Singleton Pattern [closed]

Answer 1

It was mentioned that i should post my answer from over here.

There is a technique called "Lock-free initialization" that does what you want:

interface

function getInstance: TObject;

implementation

var
   AObject: TObject;

function getInstance: TObject;
var
   newObject: TObject;
begin
   if (AObject = nil) then
   begin
      //The object doesn't exist yet. Create one.
      newObject := TObject.Create;

      //It's possible another thread also created one.
      //Only one of us will be able to set the AObject singleton variable
      if InterlockedCompareExchangePointer(AObject, newObject, nil) <> nil then
      begin
         //The other beat us. Destroy our newly created object and use theirs.
         newObject.Free;
      end;
   end;

   Result := AObject;
end;

The use of InterlockedCompareExchangePointer erects a full memory barrier around the operation. It is possible one might be able to get away with InterlockedCompareExchangePointerAcquire or InterlockedCompareExchangeRelease to get away with an optimization by only having a memory fence before or after. The problem with that is:

• i'm not smart enough to know if Acquire or Release semantics will work
• you're constructing an object, the memory barrier performance hit is the least of your worries (it's the thread safety)

InterlockedCompareExchangePointer

Windows didn't add InterlockedCompareExchangePointer until sometime around 2003. In reality it is simply a wrapper around InterlockedCompareExchange

function InterlockedCompareExchangePointer(var Destination: Pointer; Exchange: Pointer; Comparand: Pointer): Pointer stdcall;
const
    SPointerAlignmentError = 'Parameter to InterlockedCompareExchangePointer is not 32-bit aligned';
begin
{IFDEF Debug}
    //On 64-bit systems, the pointer must be aligned to 64-bit boundaries.
    //On 32-bit systems, the pointer must be aligned to 32-bit boundaries.
    if ((NativeInt(Destination) mod 4) <> 0)
            or ((NativeInt(Exchange) mod 4) <> 0)
            or ((NativeInt(Comparand) mod 4) <> 0) then
    begin
        OutputDebugString(SPointerAlignmentError);
        if IsDebuggerPresent then
            Windows.DebugBreak;
    end;
{ENDIF}
    Result := Pointer(IntPtr(InterlockedCompareExchange(Integer(IntPtr(Destination)), IntPtr(Exchange), IntPtr(Comparand))));
end;

In XE6, i find InterlockedcompareExchangePointer implemented for 32-bit in Windows.Winapi implemented the same way (except for the safety checking):

{$IFDEF WIN32}
function InterlockedCompareExchangePointer(var Destination: Pointer; Exchange: Pointer; Comparand: Pointer): Pointer; inline;
begin
  Result := Pointer(IntPtr(InterlockedCompareExchange(Integer(IntPtr(Destination)), IntPtr(Exchange), IntPtr(Comparand))));
end;
{$ENDIF}

In newer versions of Delphi you would, ideally, use the TInterlocked helper class from System.SyncObjs:

if TInterlocked.CompareExchange({var}AObject, newObject, nil) <> nil then
begin
   //The other beat us. Destroy our newly created object and use theirs.
   newObject.Free;
end;

Note: Any code released into public domain. No attribution required.

Answer 2

I think if I wanted an object-like thing that didn't have any means of being constructed I'd probably use an interface with the implementing object contained in the implementation section of a unit.

I'd expose the interface by a global function (declared in the interface section). The instance would be tidied up in a finalization section.

To get thread-safety I'd use either a critical section (or equivalent) or possibly carefully implemented double-checked locking but recognising that naive implementations only work due to the strong nature of the x86 memory model.

It would look something like this:

unit uSingleton;

interface

uses
  SyncObjs;

type
  ISingleton = interface
    procedure DoStuff;
  end;

function Singleton: ISingleton;

implementation

type
  TSingleton = class(TInterfacedObject, ISingleton)
  private
    procedure DoStuff;
  end;

{ TSingleton }

procedure TSingleton.DoStuff;
begin
end;

var
  Lock: TCriticalSection;
  _Singleton: ISingleton;

function Singleton: ISingleton;
begin
  Lock.Acquire;
  Try
    if not Assigned(_Singleton) then
      _Singleton := TSingleton.Create;
    Result := _Singleton;
  Finally
    Lock.Release;
  End;
end;

initialization
  Lock := TCriticalSection.Create;

finalization
  Lock.Free;

end.

Answer 3

The trouble with Delphi is that you always inherit the Create constructor from TObject. But we can deal with that pretty nicely! Here's a way:

TTrueSingleton = class
private
  class var FSingle: TTrueSingleton;
  constructor MakeSingleton;
public
  constructor Create;reintroduce;deprecated 'Don''t use this!';

  class function Single: TTrueSingleton;
end;

As you can see we can have a private constructor and we can hide the inherited TObject.Create constructor! In the implementation of TTrueSingleton.Create you can raise an error (run-time block) and the deprecated keyword has the added benefit of providing compile-time error handling!

Here's the implementation part:

constructor TTrueSingleton.Create;
begin
  raise Exception.Create('Don''t call me directly!');
end;

constructor TTrueSingleton.MakeSingleton;
begin
end;

class function TTrueSingleton.Single: TTrueSingleton;
begin
  if not Assigned(FSingle) then FSingle := TTrueSingleton.MakeSingleton;
  Result := FSingle;
end;

If at compile time the compiler sees you doing this:

var X: TTrueSingleton := TTrueSingleton.Create;

it will give you the deprecated warning complete with the provided error message. If you're stubborn enough to ignore it, at run time, you'll not get an object but a raised exception.

---

Later edit to introduce thread-safety. First of all I must confess, for my own code I don't care about this kind of thread-safety. The probability for two threads accessing my singleton creator routine within such a short time frame it causes two TTrueSingleton objects to be created is so small it's simply not worth the few lines of code required.

But this answer wouldn't be complete without thread safety, so here's my take on the issue. I'll use a simple spin-lock (busy waiting) because it's efficient when no locking needs to be done; Besides, it only locks ones

For this to work an other class var needs to be added: class var FLock: Integer. The Singleton class function should look like this:

class function TTrueSingleton.Single: TTrueSingleton;
var Tmp: TTrueSingleton;
begin
  MemoryBarrier; // Make sure all CPU caches are in sync
  if not Assigned(FSingle) then
  begin
    Assert(NativeUInt(@FLock) mod 4 = 0, 'FLock needs to be alligned to 32 bits.');

    // Busy-wait lock: Not a big problem for a singleton implementation
    repeat
    until InterlockedCompareExchange(FLock, 1, 0) = 0; // if FLock=0 then FLock:=1;
    try
      if not Assigned(FSingle) then
      begin 
        Tmp := TTrueSingleton.MakeSingleton;
        MemoryBarrier; // Second barrier, make sure all CPU caches are in sync.
        FSingle := Tmp; // Make sure the object is fully created when we assign it to FSingle.
      end;
    finally FLock := 0; // Release lock
    end;
  end;
  Result := FSingle;
end;

Answer 4

There is a way to hide the inherited “Create” constructor of TObject. Although it is not possible to change the access level, it can be hidden with another public parameterless method with the same name: “Create”. This simplifies the implementation of the Singleton class tremendously. See the simplicity of the code:

unit Singleton;

interface

type
  TSingleton = class
  private
    class var _instance: TSingleton;
  public
    //Global point of access to the unique instance
    class function Create: TSingleton;

    destructor Destroy; override;
  end;

implementation

{ TSingleton }

class function TSingleton.Create: TSingleton;
begin
  if (_instance = nil) then
    _instance:= inherited Create as Self;

  result:= _instance;
end;

destructor TSingleton.Destroy;
begin
  _instance:= nil;
  inherited;
end;

end.

I added the details to my original post: http://www.yanniel.info/2010/10/singleton-pattern-delphi.html

Answer 5

The most effective way to make sure something cannot be instantiated is by making it a pure abstract class. That is, if you care enough to heed compiler hints and warnings.

Then define a function in the implementation section that returns a reference to that abstract class. Like Cosmin does in one of his answers.

The implementation section implements that function (you can even make use of lazy instantiation here, as Cosmin also shows/ed).

But the crux is to have a concrete class declared and implemented in the implementation section of the unit so only the unit can instantiated it.

interface

type
  TSingleton = class(TObject)
  public
    procedure SomeMethod; virtual; abstract;
  end;

  function Singleton: TSingleton;

implementation

var
  _InstanceLock: TCriticalSection;
  _SingletonInstance: TSingleTon;

type
  TConcreteSingleton = class(TSingleton)
  public
    procedure SomeMethod; override;
  end;

function Singleton: TSingleton;
begin
  _InstanceLock.Enter;
  try
    if not Assigned(_SingletonInstance) then
      _SingletonInstance := TConcreteSingleton.Create;

    Result := _SingletonInstance;
  finally
    _InstanceLock.Leave;
  end;
end;

procedure TConcreteSingleton.SomeMethod;
begin
  // FLock can be any synchronisation primitive you like and should of course be
  // instantiated in TConcreteSingleton's constructor and freed in its destructor.
  FLock.Enter;  
  try
  finally
    FLock.Leave;
  end;
end;

That said, please bear in mind that there are plenty of problems using singletons: http://jalf.dk/blog/2010/03/singletons-solving-problems-you-didnt-know-you-never-had-since-1995/

Thread safety

David is absolutely right in his comment that I was wrong before about the function not needing any protection. The instantiation does indeed need protecting or you could end up with two (possibly more) instances of the singleton and several of them in limbo with regard to freeing (which would be done in the finalization section as with many lazy instantion mechanisms). So here is the amended version.

To get thread safety in this setup, you need to protect the instantiation of the singleton and you need to protect all methods in the concrete class that are publicly available through its abstract ancestor. Other methods do not need to be protected as they are only be callable through the publicly available ones and so are protected by the protection in those methods.

You can protect this by a simple critical section, declared in the implementation, instantiated in the initialization and free in the finalization section. Of course the CS would have to protect the freeing of the singleton as well and should therefore be freed afterwards.

Discussing this with a colleague, we came up with a way to (mis)/(ab)use the instance pointer itself as a sort of lock mechanism. It would work, but I find it to ugly to share with the world at this point in time...

What synchronisation primitives are used to protect the publicly callable methods is entirely up to the "user" (coder) and may tailored to the purpose the singleton.

Answer 6

For threadsafety you should use a lock around the create in "TTestClass.GetInstance".

procedure CreateSingleInstance(aDestination: PPointer; aClass: TClass);
begin
  System.TMonitor.Enter(Forms.Application);
  try
    if aDestination^ = nil then  //not created in the meantime?
      aDestination^ := aClass.Create;
  finally
    System.TMonitor.Exit(Forms.Application);
  end;
end;

Threadsafe:

if not Assigned(FInstance) then
  CreateSingleInstance(@FInstance, TTestClass);      

And you could raise an exception in case someone tries to create it via the normal .Create (make a private constructor CreateSingleton)