What is the benefit of using out/ref versus returning?

Question

I'm making a game using XNA framework, so I use a lot functions that operate on vectors. (especially Vector2 (64bit struct)). What bothers me is that most of the methods are defined with ref and out parameters. Here is an example:

void Min(ref Vector2 value1, ref Vector2 value2, out Vector2 result)

which looks a bit strange too me. There is also another Min which is more obvious

public static Vector2 Min(Vector2 value1, Vector2 value2);

Basically, almost all the functions have overloads with refs and outs. Similar, other APIs.

What is the benefit of this design? XNA is optimized for performance, could it be a result? Say, Quaternion requires 128b where passing by ref less.

EDIT:

Here is a test code:

public class Game1 : Microsoft.Xna.Framework.Game
{
    GraphicsDeviceManager graphics;
    SpriteBatch spriteBatch;

    private Vector2 vec1 = new Vector2(1, 2);
    private Vector2 vec2 = new Vector2(2, 3);
    private Vector2 min;
    private string timeRefOut1;
    private string timeRefOut2;
    private SpriteFont font;

    public Game1()
    {
        graphics = new GraphicsDeviceManager(this);
        Content.RootDirectory = "Content";

        refOut1();
        refOut2();
    }

    private Vector2 refOut1()
    {
        Vector2 min = Vector2.Min(vec1, vec2);
        return min;
    }

    private Vector2 refOut2()
    {
        Vector2.Min(ref vec1, ref vec2, out min);
        return min;
    }

    protected override void Initialize()
    {
        const int len = 100000000;
        Stopwatch stopWatch = new Stopwatch();
        stopWatch.Start();
        for (int i = 0; i < len; i++)
        {
            refOut1();
        }
        stopWatch.Stop();

        timeRefOut1 = stopWatch.ElapsedMilliseconds.ToString();

        stopWatch.Reset();
        stopWatch.Start();
        for (int i = 0; i < len; i++)
        {
            refOut2();
        }
        stopWatch.Stop();

        timeRefOut2 = stopWatch.ElapsedMilliseconds.ToString();

        base.Initialize();
    }

    protected override void LoadContent()
    {
        spriteBatch = new SpriteBatch(GraphicsDevice);
        font = Content.Load<SpriteFont>("SpriteFont1");
    }

    protected override void Update(GameTime gameTime)
    {
        if (GamePad.GetState(PlayerIndex.One).Buttons.Back == ButtonState.Pressed)
            this.Exit();

        base.Update(gameTime);
    }

    protected override void Draw(GameTime gameTime)
    {
        GraphicsDevice.Clear(Color.CornflowerBlue);

        spriteBatch.Begin();
        spriteBatch.DrawString(font, timeRefOut1, new Vector2(200, 200), Color.White);
        spriteBatch.DrawString(font, timeRefOut2, new Vector2(200, 300), Color.White);
        spriteBatch.End();

        // TODO: Add your drawing code here

        base.Draw(gameTime);
    }
}

The results:

• refOut1 2200
• refOut2 1400

Win 7 64bit, .Net 4. XNA 4.0

Also IL code

.method public hidebysig static void  Min(valuetype Microsoft.Xna.Framework.Vector2& value1,
                                          valuetype Microsoft.Xna.Framework.Vector2& value2,
                                          [out] valuetype Microsoft.Xna.Framework.Vector2& result) cil managed
{
  // Code size       69 (0x45)
  .maxstack  3
  IL_0000:  ldarg.2
  IL_0001:  ldarg.0
  IL_0002:  ldfld      float32 Microsoft.Xna.Framework.Vector2::X
  IL_0007:  ldarg.1
  IL_0008:  ldfld      float32 Microsoft.Xna.Framework.Vector2::X
  IL_000d:  blt.s      IL_0017
  IL_000f:  ldarg.1
  IL_0010:  ldfld      float32 Microsoft.Xna.Framework.Vector2::X
  IL_0015:  br.s       IL_001d
  IL_0017:  ldarg.0
  IL_0018:  ldfld      float32 Microsoft.Xna.Framework.Vector2::X
  IL_001d:  stfld      float32 Microsoft.Xna.Framework.Vector2::X
  IL_0022:  ldarg.2
  IL_0023:  ldarg.0
  IL_0024:  ldfld      float32 Microsoft.Xna.Framework.Vector2::Y
  IL_0029:  ldarg.1
  IL_002a:  ldfld      float32 Microsoft.Xna.Framework.Vector2::Y
  IL_002f:  blt.s      IL_0039
  IL_0031:  ldarg.1
  IL_0032:  ldfld      float32 Microsoft.Xna.Framework.Vector2::Y
  IL_0037:  br.s       IL_003f
  IL_0039:  ldarg.0
  IL_003a:  ldfld      float32 Microsoft.Xna.Framework.Vector2::Y
  IL_003f:  stfld      float32 Microsoft.Xna.Framework.Vector2::Y
  IL_0044:  ret
} // end of method Vector2::Min

and

.method public hidebysig static valuetype Microsoft.Xna.Framework.Vector2 
        Min(valuetype Microsoft.Xna.Framework.Vector2 value1,
            valuetype Microsoft.Xna.Framework.Vector2 value2) cil managed
{
  // Code size       80 (0x50)
  .maxstack  3
  .locals init (valuetype Microsoft.Xna.Framework.Vector2 V_0)
  IL_0000:  ldloca.s   V_0
  IL_0002:  ldarga.s   value1
  IL_0004:  ldfld      float32 Microsoft.Xna.Framework.Vector2::X
  IL_0009:  ldarga.s   value2
  IL_000b:  ldfld      float32 Microsoft.Xna.Framework.Vector2::X
  IL_0010:  blt.s      IL_001b
  IL_0012:  ldarga.s   value2
  IL_0014:  ldfld      float32 Microsoft.Xna.Framework.Vector2::X
  IL_0019:  br.s       IL_0022
  IL_001b:  ldarga.s   value1
  IL_001d:  ldfld      float32 Microsoft.Xna.Framework.Vector2::X
  IL_0022:  stfld      float32 Microsoft.Xna.Framework.Vector2::X
  IL_0027:  ldloca.s   V_0
  IL_0029:  ldarga.s   value1
  IL_002b:  ldfld      float32 Microsoft.Xna.Framework.Vector2::Y
  IL_0030:  ldarga.s   value2
  IL_0032:  ldfld      float32 Microsoft.Xna.Framework.Vector2::Y
  IL_0037:  blt.s      IL_0042
  IL_0039:  ldarga.s   value2
  IL_003b:  ldfld      float32 Microsoft.Xna.Framework.Vector2::Y
  IL_0040:  br.s       IL_0049
  IL_0042:  ldarga.s   value1
  IL_0044:  ldfld      float32 Microsoft.Xna.Framework.Vector2::Y
  IL_0049:  stfld      float32 Microsoft.Xna.Framework.Vector2::Y
  IL_004e:  ldloc.0
  IL_004f:  ret
} // end of method Vector2::Min

Seems overhead is caused by temp Vector. Also I tried 1GHz WP 7.5 device:

• 1979
• 1677

Number of ticks for an order of magnitude smaller number of iterations.

Answer 1

Vector2 is a struct, which means that when it's returned as a value a copy is returned, rather than returning a reference to an existing structure. By using ref/out parameters you can avoid this copy so that the Vector created in the Min method is the exact vector in your result variable.

It's one of those micro optimization that normally would be discouraged, but in the game world it's done often enough, and in environments where performance matters enough, that it's worth the slightly less readable option.