Explicit and implicit type casting is a common programming topic for almost any imperative programming language. Most C, C++, or Pascal programmers care about efficiency and speed of their code; but those who use managed programming environments, such as Java, Visual Basic, or C# rely all the optimizing tasks on the compiler and the runtime environment.
This can be a good approach in many cases, but managed languages are becoming more and more popular also for high-performance applications where the knowledge of the language, compiler, and runtime environment can enhance a program's quality and speed.
This article analyzes the most common type casting situations and the compiler behavior in them. We are going to study the MSIL generated code, but not the machine-specific instruction sequences due to the implementation and vendor dependency.
Casting primitive types
Primitive types are those non-composed types which can be handled directly by the (virtual) machine instructions, i.e.,
float, etc... Those types doesn't have inner structure, and are always passed by value if the programmer doesn't specify explicitly other behavior (using the
ref modifiers). Let's see a simple example about using and casting primitive types:
int z = 10;
double r = 3.4;
uint n = 20;
r = z; z = (int)r; n = (uint)z;
This sample performs some conversions in the set of primitive types, leaving in some cases the casting tasks to the compiler and marking conversions explicitly in some other cases.
OK, time to dive into the MSIL generated code and check the impact of type casts in our code:
.locals init ( int32 z,
 float32 r,
 unsigned int32 n)
IL_0000: ldc.i4.s 10
IL_0003: ldc.r4 (9A 99 59 40)
IL_0009: ldc.i4.s 20
IL_000b: stloc.2 IL_000c: ldloc.0
IL_000f: ldloc.1 IL_0010: conv.i4
IL_0012: ldloc.0 IL_0013: stloc.2
As we can see, there are several
Conv.XY instructions in the code, whose function is to convert the value at the top of the stack to the type designed in the opcode (r4, i4, etc...). From now, we know that the "innocent" explicit and implicit conversions between primitive types generate instructions which can be avoided with a consistent type usage. The same conversions are applied in 64-bit data types, such as
Note that the last type cast doesn't need an explicit "
Conv" opcode due to the nature of the involved types:
uint; these types have a very close storage structure (big endian bit order with a sign bit in the signed type) and conversion sign issues must be controlled by the programmer.
A special kind of primitive type is
bool (handled internally as an
int), whose conversions to numeric types (and backward) are not allowed in C#, so we will not study them.
Downcasting object references
C# provides two ways for casting object references (note that all types, unless those studied in the previous section, are reference types):
object myClass = new MyClass();
((MyClass)myClass).DoSome(); (myClass as MyClass).DoSome();
The previous is a good example of downcasting (casting from the top to the bottom of the class hierarchy). The method used to perform the cast appears to be the same, but the generated MSIL sequences are a bit different:
.locals init ( object myClass)
IL_0000: newobj instance void Sample.MyClass::.ctor()
IL_0006: ldloc.0 IL_0007: castclass Sample.MyClass
IL_000c: callvirt instance void Sample.MyClass::DoSome()
IL_0011: ldloc.0 IL_0012: isinst Sample.MyClass
IL_0017: callvirt instance void Sample.MyClass::DoSome()
In the first line of code, the compiler emits a "
Castclass" opcode, which converts the reference to the type specified between the parenthesis if possible (if not, an
InvalidCastException exception is thrown).
In the second case, the
as operator is translated as an "
IsInst" opcode, which works much faster, because it only checks the reference type but doesn't perform any sort of cast (nor throws any exception).
In performance terms, we prefer the second option, because the "
IsInst" speeds up much more the code execution, avoiding type casts and exception throwing. Here is a sample of the speed increment obtained using the "
In the other hand, parenthesized casts give a better error control to programmers, avoiding the null-reference errors obtained when invalid typecasts happen using the "
Upcasting object references
Let's make the opposite! Now it's time for climbing up into the class hierarchy, and see how slow (or fast) are these sort of casts. The following example creates an object of the type
MyDerivedClass and stores its reference in a
MyClass type variable:
MyDerivedClass myDerivedClass = new MyDerivedClass();
MyClass myClass = myDerivedClass;
And the produced code is:
.locals init ( class Sample.MyDerivedClass myDerivedClass,
 class Sample.MyClass myClass)
IL_0000: newobj instance void Sample.MyDerivedClass::.ctor()
As we can see, there are no conversion opcodes, just reference loading and storing. This is good for out efficiency purposes... as expected, upcasting type checks are made at compile time and the runtime costs are as cheap as a simple assign between variables of the same type.
C# language contains a great feature which allows to define implicit and explicit conversion operators. The efficiency of these casting methods depends on the casting method implementation. Anyway, these functions are always static and have only one parameter, so the procedure call overhead is small (no "
this" parameter should be passed). Anyway, it seems to be that the Microsoft C# compiler doesn't inline those methods, so arranging parameters and return addresses in the stack may slow your code execution speed.
Putting it all together
Here are some general tips for optimizing your programs based on the results obtained in the previous sections:
- Numeric type conversions are usually expensive, take them out of the loops and recursive functions and use the same numeric types when possible.
- Downcasting is a great invention but the type checks involved have a great impact on execution performance, check the object types out of loops and recursive functions, and use the "
as" operator into them.
- Upcasting is cheap!, use it everywhere you need.
- Build lightweight conversion operators to make custom casts faster.
All the tests and disassemblies have been made using the tools included in the .NET Framework SDK. ILDasm can tell you much about your program's performance flaws, so play with it.
Emilio is a Computer Engineer currently working as software engineer in embedded systems.
Main interests are C/C++ programming, algorithmics, compilers, embedded systems, cryptography, and operating systems.