Why is OfType<> faster than Cast<>?

Question

In answer to the following question: How to convert MatchCollection to string array

Given The two Linq expressions:

var arr = Regex.Matches(strText,          


        

Answer 1

My benchmarking does not agree with your benchmarking.

I ran an identical benchmark to Alex's and got the opposite result. I then tweaked the benchmark somewhat and again observed Cast being faster than OfType.

There's not much in it, but I believe that Cast does have the edge, as it should because its iterator is simpler. (No is check.)

Edit: Actually after some further tweaking I managed to get Cast to be 50x faster than OfType.

Below is the code of the benchmark that gives the biggest discrepancy I've found so far:

Stopwatch sw1 = new Stopwatch();
Stopwatch sw2 = new Stopwatch();

var ma = Enumerable.Range(1, 100000).Select(i => i.ToString()).ToArray();

var x = ma.OfType().ToArray();
var y = ma.Cast().ToArray();

for (int i = 0; i < 1000; i++)
{
    if (i%2 == 0)
    {
        sw1.Start();
        var arr = ma.OfType().ToArray();
        sw1.Stop();
        sw2.Start();
        var arr2 = ma.Cast().ToArray();
        sw2.Stop();
    }
    else
    {
        sw2.Start();
        var arr2 = ma.Cast().ToArray();
        sw2.Stop();
        sw1.Start();
        var arr = ma.OfType().ToArray();
        sw1.Stop();
    }
}
Console.WriteLine("OfType: " + sw1.ElapsedMilliseconds.ToString());
Console.WriteLine("Cast: " + sw2.ElapsedMilliseconds.ToString());
Console.ReadLine();

Tweaks I've made:

• Perform the "generate a list of strings" work once, at the start, and "crystallize" it.
• Perform one of each operation before starting timing - I'm not sure if this is necessary but I think it means the JITter generates code beforehand rather than while we're timing?
• Perform each operation multiple times, not just once.
• Alternate the order in case this makes a difference.

On my machine this results in ~350ms for Cast and ~18000ms for OfType.

I think the biggest difference is that we're no longer timing how long MatchCollection takes to find the next match. (Or, in my code, how long int.ToString() takes.) This drastically reduces the signal-to-noise ratio.

Edit: As sixlettervariables pointed out, the reason for this massive difference is that Cast will short-circuit and not bother casting individual items if it can cast the whole IEnumerable. When I switched from using Regex.Matches to an array to avoid measuring the regex processing time, I also switched to using something castable to IEnumerable and thus activated this short-circuiting. When I altered my benchmark to disable this short-circuiting, I get a slight advantage to Cast rather than a massive one.