What is the difference between ldobj and ldind.<type>, and why is ldobj faster?

Question

When using a 64-bit sized struct, the following code snippet

[StructLayout(LayoutKind.Explicit, Pack = 1, Size = 8)]
  unsafe struct BUF
  {
  }

((BUF*)dst) = *((BUF*)src); 

Produces

    IL_0046: nop          
    IL_0047: ldloc.s dst                              
    IL_0049: ldloc.2                                       
    IL_004a: ldobj MyClass/BUF           
    IL_004f: stobj MyClass/BUF

However, when just using a long, the following code produces

*((long*)dst) = *((long*)src); 

Produces:

IL_0046: nop
IL_0047: ldloc.s dst
IL_0049: ldloc.2
IL_004a: ldind.i8                       
IL_004b: stind.i8         

Does anyone know what difference ldobj/stobj and ldind.i8/stind.i8 makes, if any, for this example?

ldobj/stobj seems to give a 20% performance improvement, but I cannot figure out why that is happening. Aren't these two lines doing the exact same thing?

Thanks!

edit: [64-bit release mode] The bytecode looks the same when compiled in release mode. The performance measurement was done a while ago in release mode.

Answer 1

I've replicated the two distinct methods you're using, and see the same IL being generated, however the jitted code for the two is absolutely identical when run in Release mode:

Here's the test method I used:

using System;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;

namespace Test
{
    [StructLayout(LayoutKind.Explicit, Pack = 1, Size = 8)]
    unsafe struct BUF
    {
    }

    static class Program
    {
        static void Main()
        {
            BUF x, y, z;
            unsafe
            {
                Do1(&x, &y);
                Do2(&y, &z);
            }
            // Readline here to allow attaching debugger and dumping jitted code
            Console.ReadLine();
        }

        // Disable inlining to permit easier identification of the code
        [MethodImpl(MethodImplOptions.NoInlining)]
        unsafe static void Do1(BUF* src, BUF* dst)
        {
            *((BUF*)dst) = *((BUF*)src);
        }

        // Disable inlining to permit easier identification of the code
        [MethodImpl(MethodImplOptions.NoInlining)]
        unsafe static void Do2(BUF* src, BUF* dst)
        {
            *((long*)dst) = *((long*)src);
        }
    }
}

The IL for the two methods match yours:

Do1:

IL_0000: ldarg.1 
IL_0001: ldarg.0 
IL_0002: ldobj Test.BUF
IL_0007: stobj Test.BUF
IL_000c: ret 

Do2:

IL_0000: ldarg.1 
IL_0001: ldarg.0 
IL_0002: ldind.i8 
IL_0003: stind.i8 
IL_0004: ret 

And dumping the jitted code:

Do1:

Test.Program.Do1(Test.BUF*, Test.BUF*)
Begin 000007ff00170190, size 7
000007ff`00170190 488b01          mov     rax,qword ptr [rcx]
000007ff`00170193 488902          mov     qword ptr [rdx],rax
000007ff`00170196 c3              ret

Do2:

Test.Program.Do2(Test.BUF*, Test.BUF*)
Begin 000007ff001701b0, size 7
000007ff`001701b0 488b01          mov     rax,qword ptr [rcx]
000007ff`001701b3 488902          mov     qword ptr [rdx],rax
000007ff`001701b6 c3              ret

They look pretty identical to me.