Is foreach around 9% slower than old-school for-loop?

Thanks to #covid19, I've been trying to find ways to kill the lock-down boredom. I decided to investigate the performance of the traditional "for" loop compared to the somewhat more modern (well, modern in the last 20 years, haha) "foreach" loop.

I remember back in the day when I was actually writing critical-path production code when a co-worker told me: don't use foreach, always use old-school for-loop because the former is much slower

The hypothesis is that the foreach requires use of reflection by the compiler to determine the exact data type of the element in the foreach, and that extra computation results in a slightly less optimized iteration than the traditional loop. True? Well...

I ran the code below for 100s of time. The result? Yeah, anywhere between 8.5% and 10% in favor of the old-school for-loop. Guess my co-worker was right... Cheers, stay safe, love, ACC.

for each: 35720075
old-school for: 32380036
old-school is 9.35059346879871% faster than foreach

public void TestFors()
{
Random rd = new Random();
string test = new string((char)(rd.Next(0, 26) + 'a'), 1000000000);

char t = ' ';
long ticksBefore = DateTime.Now.Ticks;
foreach (char c in test) t = c;
long tickAfter = DateTime.Now.Ticks;
long x = tickAfter - ticksBefore;

Console.WriteLine("for each: {0}", x);

t = ' ';
ticksBefore = DateTime.Now.Ticks;
for (int i = 0; i < test.Length; i++) t = test[i];
tickAfter = DateTime.Now.Ticks;
long y = tickAfter - ticksBefore;

Console.WriteLine("old-school for: {0}", y);

Console.WriteLine("old-school is {0}% faster than foreach", 100.0 * (x - y) / x);
}

Comments

  1. Not sure if I agree with the methodology :-)

    Here are my results using a more pragmatic approach(*):

    End of warm-up phase: 0
    End of collection phase: 0
    For loop
    Average: 20781.5728
    P50: 20000
    P90: 20035
    P95: 29997
    P99: 30025

    End of warm-up phase: 0
    End of collection phase: 0
    Foreach loop
    Average: 20551.2984
    P50: 20000
    P90: 20033
    P95: 20211
    P99: 30028

    Based on that, I'd say the difference is insignificant.

    *Approach:

    1. Instead of a single 1 billion entry loop (is that common?) I did 1000 loops over 1 million entries.

    2. Full GC before running each test (for or foreach).

    3. Ran 1000 loops for warm-up caches before running 1000 loops collecting time information.

    Reply Delete
    Replies

    1. Source code:

      class PerformanceTests
      {
      private static Random random = new Random(0);
      private static int[] randomData = BuildRandomData(1000000);

      private static int[] BuildRandomData(int length)
      {
      int[] result = new int[length];

      for (int i = 0; i < length; i++)
      {
      result[i] = random.Next();
      }

      return result;
      }

      private static void FullGC()
      {
      GC.Collect(GC.MaxGeneration, GCCollectionMode.Forced, true, true);
      GC.WaitForPendingFinalizers();
      }

      private static void WriteReport(string context, long[] ticks)
      {
      Console.WriteLine(context);
      double[] durations = new double[ticks.Length - 1];
      for (int i = 0; i < durations.Length; i++)
      {
      durations[i] = ticks[i + 1] - ticks[i];
      }

      Array.Sort(durations);

      Console.WriteLine("Average: " + durations.Average());
      Console.WriteLine("P50: " + durations[(int)(durations.Length * .50)]);
      Console.WriteLine("P90: " + durations[(int)(durations.Length * .90)]);
      Console.WriteLine("P95: " + durations[(int)(durations.Length * .95)]);
      Console.WriteLine("P99: " + durations[(int)(durations.Length * .99)]);
      }

      public static void ForLoop(int iterations)
      {
      var ticks = new long[iterations + 1];

      int limit = randomData.Length;
      int dummy = 0;

      for (int it = 0; it < iterations; it++)
      {
      for (int i = 0; i < limit; i++)
      {
      dummy = dummy ^ randomData[i];
      }
      }

      Console.WriteLine("End of warm-up phase: " + dummy);

      ticks[0] = DateTime.UtcNow.Ticks;
      iterations++;

      for (int it = 1; it < iterations; it++)
      {
      for (int i = 0; i < limit; i++)
      {
      dummy = dummy ^ randomData[i];
      }

      ticks[it] = DateTime.UtcNow.Ticks;
      }

      Console.WriteLine("End of collection phase: " + dummy);

      WriteReport("For loop", ticks);
      }

      public static void ForeachLoop(int iterations)
      {
      var ticks = new long[iterations + 1];

      int limit = randomData.Length;
      int dummy = 0;

      for (int it = 0; it < iterations; it++)
      {
      foreach (int value in randomData)
      {
      dummy = dummy ^ value;
      }
      }

      Console.WriteLine("End of warm-up phase: " + dummy);

      ticks[0] = DateTime.UtcNow.Ticks;
      iterations++;

      for (int it = 1; it < iterations; it++)
      {
      foreach (int value in randomData)
      {
      dummy = dummy ^ value;
      }

      ticks[it] = DateTime.UtcNow.Ticks;
      }

      Console.WriteLine("End of collection phase: " + dummy);

      WriteReport("Foreach loop", ticks);
      }

      public static void Run()
      {
      FullGC();
      ForLoop(10000);
      FullGC();
      ForeachLoop(10000);
      }
      }

      Delete

    2. Sorry, I ran 10,000 iterations over 1,000,000 elements, not 1,000 iterations.

      Delete

Post a Comment

[フレーム]

Popular posts from this blog

Quasi FSM (Finite State Machine) problem + Vibe

Not really an FSM problem since the state isn't changing, it is just defined by the current input. Simply following the instructions should do it. Using VSCode IDE you can also engage the help of Cline or Copilot for a combo of coding and vibe coding, see below screenshot. Cheers, ACC. Process String with Special Operations I - LeetCode You are given a string  s  consisting of lowercase English letters and the special characters:  * ,  # , and  % . Build a new string  result  by processing  s  according to the following rules from left to right: If the letter is a  lowercase  English letter append it to  result . A  '*'   removes  the last character from  result , if it exists. A  '#'   duplicates  the current  result  and  appends  it to itself. A  '%'   reverses  the current  result . Return the final string  result  after processing all char...

Shortest Bridge – A BFS Story (with a Twist)

Here's another one from the Google 30 Days challenge on LeetCode — 934. Shortest Bridge . The goal? Given a 2D binary grid where two islands (groups of 1s) are separated by water (0s), flip the fewest number of 0s to 1s to connect them. Easy to describe. Sneaky to implement well. 🧭 My Approach My solution follows a two-phase Breadth-First Search (BFS) strategy: Find and mark one island : I start by scanning the grid until I find the first 1 , then use BFS to mark all connected land cells as 2 . I store their positions for later use. Bridge-building BFS : For each cell in the marked island, I run a BFS looking for the second island. Each BFS stops as soon as it hits a cell with value 1 . The minimum distance across all these searches gives the shortest bridge. πŸ” Code Snippet Here's the core logic simplified: public int ShortestBridge(int[][] grid) { // 1. Mark one island as '2' and gather its coordinates List<int> island = FindAndMark...

Classic Dynamic Programming IX

A bit of vibe code together with OpenAI O3. I asked O3 to just generate the sieve due to laziness. Sieve is used to calculate the first M primes (when I was using Miller-Rabin, was giving me TLE). The DP follows from that in a straightforward way: calculate the numbers from i..n-1, then n follows by calculating the min over all M primes. Notice that I made use of Goldbach's Conjecture as a way to optimize the code too. Goldbach's Conjecture estates that any even number greater than 2 is the sum of 2 primes. The conjecture is applied in the highlighted line. Cheers, ACC. PS: the prompt for the sieve was the following, again using Open AI O3 Advanced Reasoning: " give me a sieve to find the first M prime numbers in C#. The code should produce a List<int> with the first M primes " Minimum Number of Primes to Sum to Target - LeetCode You are given two integers  n  and  m . You have to select a multiset of  prime numbers  from the  first   m  pri...