What is the cost of casting in .NET (C#)?
As I was (at that moment) running out of ideas about optimizations in hot paths in FbNetExternalEngine, I turned my attention into other parts of code. And one piece that was doing casting on every call caught my attention. Could I make the casting faster? What is actually the cost of casting? Does it matter? I had to dive into it.
Setup
The setup of this benchmark was pretty straightforward. In my case the casting was from object to ITuple. Your mileage may vary.
BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19043.1466 (21H1/May2021Update)
AMD Ryzen 7 5800X, 1 CPU, 16 logical and 8 physical cores
.NET SDK=6.0.101
[Host] : .NET 6.0.1 (6.0.121.56705), X64 RyuJIT
DefaultJob : .NET 6.0.1 (6.0.121.56705), X64 RyuJIT
Method argument
First, I tried casting the argument of a method.
public ITuple UnsafeCastArgument(object o)
{
return Unsafe.As<ITuple>(o);
}
public ITuple RegularCastArgument(object o)
{
return (ITuple)o;
}
| Method | o | Mean | Error | StdDev | Ratio | Code Size |
|---|---|---|---|---|---|---|
| UnsafeCastArgument | ? | 0.0002 ns | 0.0003 ns | 0.0003 ns | 0.000 | 4 B |
| RegularCastArgument | ? | 1.0396 ns | 0.0044 ns | 0.0036 ns | 1.000 | 25 B |
| UnsafeCastArgument | (10) | 0.0003 ns | 0.0005 ns | 0.0005 ns | 0.000 | 4 B |
| RegularCastArgument | (10) | 2.1023 ns | 0.0044 ns | 0.0041 ns | 1.000 | 25 B |
| UnsafeCastArgument | (10, 20) | 0.0001 ns | 0.0002 ns | 0.0002 ns | 0.000 | 4 B |
| RegularCastArgument | (10, 20) | 2.1028 ns | 0.0056 ns | 0.0049 ns | 1.000 | 25 B |
| UnsafeCastArgument | (10, (…)8, 9) [38] | 0.0002 ns | 0.0005 ns | 0.0005 ns | 0.000 | 4 B |
| RegularCastArgument | (10, (…)8, 9) [38] | 2.3068 ns | 0.0033 ns | 0.0027 ns | 1.000 | 25 B |
The ? denotes null, rest is regular ValueTuples of various sizes.
The unsafe casting is, as expected, a clear winner. Looking at the disassembly makes it obvious.
mov rax,rdx
ret
Could this be any smoother? 😃
The safe variant looks like this.
sub rsp,28
mov rcx,offset MT_System.Runtime.CompilerServices.ITuple
call CORINFO_HELP_CHKCASTINTERFACE
nop
add rsp,28
ret
If you’d like to check implementation of CORINFO_HELP_CHKCASTINTERFACE, you can start here.
Property
For the completeness-sake I also tested how the code behaves when casting property.
public object O { get; set; }
public ITuple UnsafeCastProperty()
{
return Unsafe.As<ITuple>(O);
}
public ITuple RegularCastProperty()
{
return (ITuple)O;
}
| Method | O | Mean | Error | StdDev | Ratio | Code Size |
|---|---|---|---|---|---|---|
| UnsafeCastProperty | ? | 0.0009 ns | 0.0008 ns | 0.0007 ns | 0.001 | 5 B |
| RegularCastProperty | ? | 1.2440 ns | 0.0017 ns | 0.0014 ns | 1.000 | 29 B |
| UnsafeCastProperty | (10) | 0.0007 ns | 0.0008 ns | 0.0007 ns | 0.000 | 5 B |
| RegularCastProperty | (10) | 2.3092 ns | 0.0036 ns | 0.0030 ns | 1.000 | 29 B |
| UnsafeCastProperty | (10, 20) | 0.0008 ns | 0.0013 ns | 0.0012 ns | 0.000 | 5 B |
| RegularCastProperty | (10, 20) | 2.3122 ns | 0.0024 ns | 0.0022 ns | 1.000 | 29 B |
| UnsafeCastProperty | (10, (…)8, 9) [38] | 0.0007 ns | 0.0008 ns | 0.0006 ns | 0.000 | 5 B |
| RegularCastProperty | (10, (…)8, 9) [38] | 2.1110 ns | 0.0080 ns | 0.0075 ns | 1.000 | 29 B |
The numbers are more or less the same.
The assembly is virtually identical, only difference being dealing with accessing the value from [rcx+8] instead of directly from rdx.
mov rax,[rcx+8]
ret
sub rsp,28
mov rdx,[rcx+8]
mov rcx,offset MT_System.Runtime.CompilerServices.ITuple
call CORINFO_HELP_CHKCASTINTERFACE
nop
add rsp,28
ret
Summary
Do 2 nanoseconds matter? Most of the time no. But I wanted to look at this anyway.
