CppSharp Users Manual
This tool allows you to generate .NET bindings that wrap C/C++ code allowing interoperability with managed languages. This can be useful if you have an existing native codebase and want to add scripting support, or want to consume an existing native library in your managed code.
There are not many automated binding tools around, the only real alternative is SWIG. So how is it different from SWIG?
- Cleaner bindings
- No need to generate a C layer to interop with C++.
- Based on an actual C++ parser (Clang) so very accurate.
- Understands C++ at the ABI (application binary interface) level
- Supports virtual method overriding
- Easily extensible semantics via user passes
- Strongly-typed customization APIs
- Can be used as a library
The backend of the bindings generator is abstracted and can support several different targets.
It can be changed by using the Options.GeneratorKind
option.
For .NET we support these bindings technologies:
- C# (P/Invoke)
- C++/CLI
There is also experimental support for these JavaScript-related targets:
- N-API (Node.js)
- QuickJS
- TypeScript
- Emscripten
The parser supports several different targets and needs to be correctly configured to match the compiled native code.
This can be done by using the ParserOptions.TargetTriple
option.
These triples follow the same format as LLVM/Clang, which is documented here.
Here are a few examples for the most common variants:
i686-pc-win32-msvc
x86_64-pc-win32-msvc
i686-linux-gnu
x86_64-linux-gnu
i686-apple-darwin
x86_64-apple-darwin
In this section we will go through how the generator deals C/C++ types.
char
→byte
/System::Byte
bool
→bool
/System::Boolean
short
→short
/System::Int16
int
,long
→int
/System::Int32
long long
→long
/System::Int64
Note: Signedness is also preserved in the conversions.
These size of these types are dependent on environment and compiler, so the mappings above are only representative of some environments (the specific data model is usually abbreviated as LP32, ILP32, LLP64, LP64).
Please check the fundamental types properties table at cppreference.com for more information about this.
float
→float
/System::Single
double
→double
/System::Double
wchar_t
→char
/System::Char
void
→void
/System::Void
These are mapped to .NET CLR arrays.
These are mapped to .NET CLR delegates.
This is implemented by the DelegatesPass
pass.
These are mapped to .NET CLR references unless:
void*
→System.IntPtr
/System::IntPtr
const char*
→string
/System::String
Regular non-wide strings are assumed to be ASCII by default (marshaled with .NET Marshal.PtrToStringAnsi
).
Wide strings are marshaled either as UTF-16 or UTF-32, depending on the width of wchar_t
for the target.
This behavior can be overriden by using the Options.Encoding
setting.
References are mapped to .NET CLR references just like pointers.
We do not preserve type definitions since .NET and its main language C# do not have the concept of type aliases like C/C++. There is an exception in the case of a typedef'd function (pointer) declaration. In this case we generate a .NET delegate with the name of the typedef.
Regular C/C++ enums are translated to .NET enumerations.
C and C++ enums do not introduce their own scope (different from C++11 strongly typed enums). This means the enumerated values will leak into an outer context, like a class or a namespace. When this is detected, the generator tries to map to an outer enclosing context and generate a new name.
Some enumerations represent bitfield patterns. The generator tries to check for
this with some heuristics. If there are enough values in the enum to make a good
guess, we apply the [Flags]
.NET attribute to the wrapper enum.
This is implemented by the CheckFlagEnumsPass
pass.
Since global scope functions are not supported in C# (though they are available in the CLR) they are mapped as a static function in a class, to be consumable by any CLS-compliant language.
By default all globals functions of a translation unit are mapped to a static class with the name of of the unit prefixed by the namespace.
We also provide special passes that try to move these free functions either as instance or static functions of some class.
See the FunctionToInstanceMethodPass
and FunctionToStaticMethodPass
passes.
C/C++ variadic arguments need careful handling because they are not constrained to be of the same type. .NET provides two types of variadic arguments support:
This is the preferred and idiomatic method but can only be used when we know the variadic arguments will all be of the same type. Since we have no way to derive this fact from the information in C/C++ function signatures, you will need set this explicitly.
This is a lesser known method for variadic arguments in .NET and was added by Microsoft for better C++ compatibility in the runtime. As you can guess, this does support different types per variable argument but is more verbose and less idiomatic to use. By default we use this to wrap variadic functions.
A subset of default arguments values are supported.
This is implemented by the HandleDefaultParamValuesPass
pass.
In C++, both classes and structs are identical and can be used in both heap
(malloc
/ new
) and automatic (stack) allocations.
This is unlike .NET, in which there is an explicit differentiation of the allocation semantics of the type in the form of classes (reference types) and structs (value types),
By default, classes and structs are wrapped as .NET reference types. You can provide an explicit mapping to wrap any type as a value type.
TODO: If the native type is a POD type, that means we can safely convert it to a value type. This would make the generator do the right thing by default and is pretty easy to implement.
Classes that respect the following constraints are bound as static managed classes.
- Do not provide any non-private constructors
- Do not provide non-static fields or methods
- Do not provide any static function that return a pointer to the class
This is implemented by the CheckStaticClass
pass.
Constructors are mapped to .NET class constructors.
Note: An extra constructor is generated that takes a native pointer to the class. This allows construction of managed instances from native instances.
Additionally we will create C# conversion operators out of compatible single-argument constructors.
This is implemented by the ConstructorToConversionOperatorPass
pass.
Destructors are mapped to the Dispose() pattern of .NET.
Most of the regular C++ operators can be mapped to .NET operator overloads.
In case an operator has no match in C# then its added as a named method with the same parameters.
This is implemented by the CheckOperatorsOverloads
pass.
C++ supports implementation inheritance of multiple types. This is incompatible with .NET which supports only single implementation inheritance (but multiple interface inheritance).
This is the simplest case and we can map the inheritance directly.
In this case we can only map one class directly. The others can be mapped as interfaces if they only provide pure virtual methods. Otherwise the best we can do is provide some conversion operators in .NET to get access to them.
Overriding virtual methods from managed classes is supported.
Instances of these types can be passed to native code and and whenever the native code calls one of those functions there will be a transition to the C# code.
This is done by mirroring the virtual methods table with our own table at runtime, and replacing the table entries with unmanaged function pointers that transition to managed code as needed.
Class instance fields are translated to managed properties.
This is implemented by the FieldToPropertyPass
pass.
Template parsing is supported and you can type map them to other types.
Code generation for templates is experimental and can be enabled by the GenerateClassTemplates
option.
Since C preprocessor definitions can be used for very different purposes, we can only do so much when wrapping them to managed code.
These can be translated to proper .NET enumerations.
These can be translated to .NET static constant definitions.
This case is not supported and probably never will.
This case is not supported and probably never will.
We support a set of helper defines that can be used to annotate the native code with:
-
CS_IGNORE_FILE
(translation units)Used to ignore whole translation units.
-
CS_IGNORE
(declarations)Used to ignore declarations from being processed.
-
CS_IGNORE_GEN
(declarations)Used to ignore declaration from being generated.
-
CS_IGNORE_FILE
(translation units)Used to ignore all declarations of one header.
-
CS_VALUE_TYPE
(classes and structs)Used to flag that a class or struct is a value type.
-
CS_IN_OUT
/CS_OUT
(parameters)Used in function parameters to specify their usage kind.
-
CS_FLAGS
(enums)Used to specify that enumerations represent bitwise flags.
-
CS_READONLY
(fields and properties)Used in fields and properties to specify read-only semantics.
-
CS_EQUALS
/CS_HASHCODE
(methods)Used to flag method as representing the .NET Equals or Hashcode methods.
-
CS_CONSTRAINT(TYPE [, TYPE]*)
(templates)Used to define constraint of generated generic type or generic method.
-
CS_INTERNAL
(methods)Used to flag a method as internal to an assembly. So, it is not accessible outside that assembly.
These are implemented by the CheckMacrosPass
pass.
There is full support for parsing of Doxygen-style C++ comments syntax.
They are translated to .NET XML-style comments.
We can also figure out the intended semantic usage (ref
or out
) for parameters from Doxyxen tags.
Related passes:
Support for these features is limited:
- Exceptions
- RTTI
They are supported and taken into account by the C++ parser for bindings generation, but there is currently no way to catch C++ exceptions from C#.
There is also no way to check RTTI type information for a specific type from C#,
but not an issue in practice since C# itself provides this via GetType()
.
The generator provides some built-in type maps for the most common C/C++ standard library types:
std::string
std::wstring
(C++/CLI only (UTF-16), pending PR for C#)
These are mapped automatically to .NET strings.
std::vector
std::map
std::set
Support for wrapping these is experimental and only currently works on the CLI backend.
The generator provides various ways to customize the generation process.
If all you need to do is customize what gets generated for a type, then you can use the type maps feature. This lets you hook into the process for a specific type pattern.
If you need more control then you can write your own pass. Passes have full access to the parsed AST (Abstract Syntax Tree) so you can modify the entire structure and declaration data of the source code. This is very powerful and should allow you to pretty much do anything you want.
The generator already provides many ready-to-use passes that can transform the wrapped code to be more idiomatic:
Use these to rename your declarations automatically so they follow .NET conventions. When setting up renaming passes, you can declare what kind of declarations they apply to. There are two different kinds of rename passes:
This is a very simple to use pass that changes the case of the name of the declarations it matches.
This pass allows you to do powerful regex-based pattern matching renaming of declaration names.
This pass introduces instance methods that call a C/C++ global function. This can be useful to map "object-oriented" design in C to .NET classes. If your function takes an instance to a class type as the first argument, then you can use this pass.
This pass introduces static methods that call a C/C++ global function. This can be useful to gather related global functions inside the object it belongs to semantically.
This pass introduces a property that calls the native C/C++ getter and setter function. This can make the API much more idiomatic and easier to use under .NET languages.
Some internal functionalities are also implemented as passes like checking for invalid declaration names or resolving incomplete declarations. Please check the developer manual for more information about these.
If you're exposing C++ functions on Windows, you'll have to add the __declspec(dllexport)
directive, otherwise the symbols won't be found when calling them from the managed world. You could also add the directive to a class directly, like this:
class __declspec(dllexport) ExposedClass
{
// class definition
}