IL and Memory Allocations

.NET C#, similarly to Java, or LLVM-based languages, is a VM-based language. It gets compiled into an Intermediary Language, which is then compiled (usually with JIT) into platform-specific machine code.

IL is a stack-based language. Looking through the generated IL code, we can understand how memory is being managed in our programs. Specifically, we can see how many memory allocations there are. This is quite important since lots of allocations might cause GC to hurt our app’s performance.

Below are a few examples of C# - IL snippets with notes on interesting parts. I disabled nullability for the IL snippets to be shorter.

IL Examples

Simple Class With One Method

public class MyClass
{
    public int Increment(int number) => number + 1;
}

.class public auto ansi beforefieldinit
  HelloWorld.MyClass
    extends [System.Runtime]System.Object
{

  .method public hidebysig instance int32
    Increment(
      int32 number
    ) cil managed
  {
    .maxstack 8

    // [5 41 - 5 51]
    IL_0000: ldarg.1      // number
    IL_0001: ldc.i4.1
    IL_0002: add
    IL_0003: ret

  } // end of method MyClass::Increment

  .method public hidebysig specialname rtspecialname instance void
    .ctor() cil managed
  {
    .maxstack 8

    IL_0000: ldarg.0      // this
    IL_0001: call         instance void [System.Runtime]System.Object::.ctor()
    IL_0006: nop
    IL_0007: ret

  } // end of method MyClass::.ctor
} // end of class HelloWorld.MyClass

The example is rather straightforward, here’re some things to note:

// [5 41 - 5 51] - such comments appear often in generated IL code (at least in Debug builds) to indicate line/column numbers of corresponding C# code.
Since IL is stack-based, we will see lots of loads (instructions starting with LD), which are followed by operations that act on values on the stack (like ADD). Wikipedia contains listing of all CIL instructions. Here’s a brief explanation of instructions that appear in our code example:
- LDARG.1 - loads first argument onto the stack (in this case it’s the value of number argument). LDARG.0 would represent this.
- LDC.i4.1 - pushes “1” on top of the stack (since we’re incrementing the provided number)
- ADD - without any surprise, it adds two last values on the stack
- RET - returns from a method
- CALL - calls a method. In this case, since our class does not inherit from anything explicitly, it implicitly inherits Object, and Object’s constructor is called within our class’s constructor. Before it’s called, LDARG.0 is used to load this onto the stack. This reference is required by the constructor.
- NOP - no operation. It’s used in Debug builds, to mark points where breakpoints could be placed even though there isn’t any actual instruction (like on opening/closing braces).

Returning a Delegate From Class’s Method

public class MyClass
{
    public int Increment(int number) => number + 1;

    public Func<int, int> GetDelegate() => Increment;
}

.class public auto ansi beforefieldinit
  HelloWorld.MyClass
    extends [System.Runtime]System.Object
{

  .method public hidebysig instance int32
    Increment(
      int32 number
    ) cil managed
  {
    .maxstack 8

    // [5 41 - 5 51]
    IL_0000: ldarg.1      // number
    IL_0001: ldc.i4.1
    IL_0002: add
    IL_0003: ret

  } // end of method MyClass::Increment

  .method public hidebysig instance class [System.Runtime]System.Func`2<int32, int32>
    GetDelegate() cil managed
  {
    .custom instance void System.Runtime.CompilerServices.NullableContextAttribute::.ctor(unsigned int8)
      = (01 00 01 00 00 ) // .....
      // unsigned int8(1) // 0x01
    .maxstack 8

    // [7 44 - 7 53]
    IL_0000: ldarg.0      // this
    IL_0001: ldftn        instance int32 HelloWorld.MyClass::Increment(int32)
    IL_0007: newobj       instance void class [System.Runtime]System.Func`2<int32, int32>::.ctor(object, native int)
    IL_000c: ret

  } // end of method MyClass::GetDelegate

  .method public hidebysig specialname rtspecialname instance void
    .ctor() cil managed
  {
    .maxstack 8

    IL_0000: ldarg.0      // this
    IL_0001: call         instance void [System.Runtime]System.Object::.ctor()
    IL_0006: nop
    IL_0007: ret

  } // end of method MyClass::.ctor
} // end of class HelloWorld.MyClass

LDFTN pushes a pointer to a function/method on a stack. It is required by Func constructor (it requires a target - class instance, and functon pointer - the function to be invoked).

Every invocation of GetDelegate causes a new isntance of Func to be allocated on the heap (NEWOBJ ...Func2<int32, int32>...). We could optimize the C# code by storing a singleton instance of Func and reusing it lated:

private Func<int, int> _func = Increment;

public Func<int, int> GetDelegate() => _func;

Returning a Delegate From Lambda (no closure)

public class MyClass
{
    public Func<int, int> GetDelegate() => x => x + 1;
}

.class public auto ansi beforefieldinit
  HelloWorld.MyClass
    extends [System.Runtime]System.Object
{

  .class nested private sealed auto ansi serializable beforefieldinit
    '<>c'
      extends [System.Runtime]System.Object
  {
    .custom instance void [System.Runtime]System.Runtime.CompilerServices.CompilerGeneratedAttribute::.ctor()
      = (01 00 00 00 )

    .field public static initonly class HelloWorld.MyClass/'<>c' '<>9'

    .field public static class [System.Runtime]System.Func`2<int32, int32> '<>9__0_0'

    .method private hidebysig static specialname rtspecialname void
      .cctor() cil managed
    {
      .maxstack 8

      IL_0000: newobj       instance void HelloWorld.MyClass/'<>c'::.ctor()
      IL_0005: stsfld       class HelloWorld.MyClass/'<>c' HelloWorld.MyClass/'<>c'::'<>9'
      IL_000a: ret

    } // end of method '<>c'::.cctor

    .method public hidebysig specialname rtspecialname instance void
      .ctor() cil managed
    {
      .maxstack 8

      IL_0000: ldarg.0      // this
      IL_0001: call         instance void [System.Runtime]System.Object::.ctor()
      IL_0006: nop
      IL_0007: ret

    } // end of method '<>c'::.ctor

    .method assembly hidebysig instance int32
      '<GetDelegate>b__0_0'(
        int32 x
      ) cil managed
    {
      .maxstack 8

      // [5 55 - 5 60]
      IL_0000: ldarg.1      // x
      IL_0001: ldc.i4.1
      IL_0002: add
      IL_0003: ret

    } // end of method '<>c'::'<GetDelegate>b__0_0'
  } // end of class '<>c'

  .method public hidebysig instance class [System.Runtime]System.Func`2<int32, int32>
    GetDelegate() cil managed
  {
    .maxstack 8

    // [5 44 - 5 60]
    IL_0000: ldsfld       class [System.Runtime]System.Func`2<int32, int32> HelloWorld.MyClass/'<>c'::'<>9__0_0'
    IL_0005: dup
    IL_0006: brtrue.s     IL_001f
    IL_0008: pop
    IL_0009: ldsfld       class HelloWorld.MyClass/'<>c' HelloWorld.MyClass/'<>c'::'<>9'
    IL_000e: ldftn        instance int32 HelloWorld.MyClass/'<>c'::'<GetDelegate>b__0_0'(int32)
    IL_0014: newobj       instance void class [System.Runtime]System.Func`2<int32, int32>::.ctor(object, native int)
    IL_0019: dup
    IL_001a: stsfld       class [System.Runtime]System.Func`2<int32, int32> HelloWorld.MyClass/'<>c'::'<>9__0_0'
    IL_001f: ret

  } // end of method MyClass::GetDelegate

  .method public hidebysig specialname rtspecialname instance void
    .ctor() cil managed
  {
    .maxstack 8

    IL_0000: ldarg.0      // this
    IL_0001: call         instance void [System.Runtime]System.Object::.ctor()
    IL_0006: nop
    IL_0007: ret

  } // end of method MyClass::.ctor
} // end of class HelloWorld.MyClass

The generated IL is considerably longer than before, even thought the C# code is actually shorter. A few new instructions popped up:

STSFLD <field> - pops value from a stack a puts it in a requested static field. In this case, a new isntance of compiler-generated class <>c is created, and then its assigned to <>9 of that class.
LDSFLD <field> - pushes value of specified static field onto the stack.
BRTRUE.S <label> - jumps to the specified label if value on stack is non-zero

Let’s try to “decompile” the IL code above back into C#, together with the compiler-generated stuff:

public class MyClass
{
    private class HiddenClass
    {
        public static HiddenClass instance;
        public static Func<int, int>? func = null;

        static HiddenClass
        {
            HiddenClass.instance = new HiddenClass();
        }

        public int Lambda()
        {
            return x + 1;
        }
    }

    public Func<int, int> GetDelegate()
    {
        if (HiddenClass.func is not null)
        {
            return HiddenClass.func;
        }

        HiddenClass.func = HiddenClass.instance.Lambda;

        return HiddenClass.func;
    }
}

The most insteresting thing in classes that contain lambdas is that compiler generates additional class for us behind the scenes, and puts lamda function’s code in that class. The generated class has a static constructor (cctor instead of usual ctor).

Interestingly, when using lambda, the compiler already includes the optimization that we had to add by ourselves before. The delegate gets instantiated just once, therefore there’s just one allocation, no matter how many times we call GetDelegate.

Returning a Delegate From Lambda (with closure)

This time, we’re going to intrdocue a closure. The sample is similar to what we looked at previously, only this time our lambda will keep the state of the method where it was created:

public class MyClass
{
    public Func<int, int> GetDelegate()
    {
        var one = 1;

        return x => x + one;
    }
}

Here’s IL:

.class public auto ansi beforefieldinit
  HelloWorld.MyClass
    extends [System.Runtime]System.Object
{

  .class nested private sealed auto ansi serializable beforefieldinit
    '<>c'
      extends [System.Runtime]System.Object
  {
    .custom instance void [System.Runtime]System.Runtime.CompilerServices.CompilerGeneratedAttribute::.ctor()
      = (01 00 00 00 )

    .field public static initonly class HelloWorld.MyClass/'<>c' '<>9'

    .field public static class [System.Runtime]System.Func`2<int32, int32> '<>9__0_0'

    .method private hidebysig static specialname rtspecialname void
      .cctor() cil managed
    {
      .maxstack 8

      IL_0000: newobj       instance void HelloWorld.MyClass/'<>c'::.ctor()
      IL_0005: stsfld       class HelloWorld.MyClass/'<>c' HelloWorld.MyClass/'<>c'::'<>9'
      IL_000a: ret

    } // end of method '<>c'::.cctor

    .method public hidebysig specialname rtspecialname instance void
      .ctor() cil managed
    {
      .maxstack 8

      IL_0000: ldarg.0      // this
      IL_0001: call         instance void [System.Runtime]System.Object::.ctor()
      IL_0006: nop
      IL_0007: ret

    } // end of method '<>c'::.ctor

    .method assembly hidebysig instance int32
      '<GetDelegate>b__0_0'(
        int32 x
      ) cil managed
    {
      .maxstack 2
      .locals init (
        [0] int32 one,
        [1] int32 V_1
      )

      // [6 5 - 6 6]
      IL_0000: nop

      // [7 9 - 7 21]
      IL_0001: ldc.i4.1
      IL_0002: stloc.0      // one

      // [9 9 - 9 24]
      IL_0003: ldarg.1      // x
      IL_0004: ldloc.0      // one
      IL_0005: add
      IL_0006: stloc.1      // V_1
      IL_0007: br.s         IL_0009

      // [10 5 - 10 6]
      IL_0009: ldloc.1      // V_1
      IL_000a: ret

    } // end of method '<>c'::'<GetDelegate>b__0_0'
  } // end of class '<>c'

  .method public hidebysig instance class [System.Runtime]System.Func`2<int32, int32>
    GetDelegate() cil managed
  {
    .maxstack 8

    // [5 44 - 10 6]
    IL_0000: ldsfld       class [System.Runtime]System.Func`2<int32, int32> HelloWorld.MyClass/'<>c'::'<>9__0_0'
    IL_0005: dup
    IL_0006: brtrue.s     IL_001f
    IL_0008: pop
    IL_0009: ldsfld       class HelloWorld.MyClass/'<>c' HelloWorld.MyClass/'<>c'::'<>9'
    IL_000e: ldftn        instance int32 HelloWorld.MyClass/'<>c'::'<GetDelegate>b__0_0'(int32)
    IL_0014: newobj       instance void class [System.Runtime]System.Func`2<int32, int32>::.ctor(object, native int)
    IL_0019: dup
    IL_001a: stsfld       class [System.Runtime]System.Func`2<int32, int32> HelloWorld.MyClass/'<>c'::'<>9__0_0'
    IL_001f: ret

  } // end of method MyClass::GetDelegate

  .method public hidebysig specialname rtspecialname instance void
    .ctor() cil managed
  {
    .maxstack 8

    IL_0000: ldarg.0      // this
    IL_0001: call         instance void [System.Runtime]System.Object::.ctor()
    IL_0006: nop
    IL_0007: ret

  } // end of method MyClass::.ctor
} // end of class HelloWorld.MyClass

The result is very similar to what we had seen before. Mostly the lamba function’s logic in the compiler-generated class is more involved, local variables are introduced.

References

← Garbage Collector

gRPC →