Build Status

Use CBinding.jl to automatically create C library bindings with Julia at runtime!

Package supports these C features:

  • fully supports C's struct, union, and enum types
  • alignment strategies
  • bit fields
  • nested types
  • anonymous types
  • type qualifiers
  • variadic functions
  • unknown-length arrays
  • inline functions (experimental opt-in)
  • typed function pointers
  • function calling conventions
  • automatic callback function pointers
  • documentation generation
  • preprocessor macros (partially supported)
  • fully supports insane C (i.e. extern struct { int i; } g[2], func();)

Read on to learn how to automatically create C library bindings, or learn how to use the generated bindings.

Create bindings with CBinding.jl

First, set up a compiler context to collect C expressions (at the module scope, or at the REPL).

julia> using CBinding

julia> c``

Notice that c`...` is a command macro (with the backticks) and is the means of specifying command line arguments to the Clang parser. Each time such a command macro is used, a new compiler context is started for the module creating it. A more real-life example might look like:

julia> libpath = find_libpath();

julia> c`-std=c99 -Wall -DGO_FAST=1 -Imylib/include -L$(libpath) -lmylib`

The compiler context also finds the paths of all specified libraries so it can use them in any bindings that are created.

Next the c"..." string macro can be used to input C code and automatically create the equivalent Julia types, global variable bindings, and function bindings. It is often the case that the C code will span multiple lines, so the triple-quoted variant (c"""...""") is most effective for this usage.

julia> c"""
         struct S;
         struct T {
           int i;
           struct S *s;
           struct T *t;
         extern void func(struct S *s, struct T t);

That's it... That's all that is needed to create a couple C types and a function binding in Julia, but actually, it gets even easier!

C API's usually come with header files, so let's just use those to create the Julia bindings and save some effort. By default, bindings are generated from the code directly written in C string macros and header files explicitly included in them, but not headers included by those headers. See the i string macro option to allow parsing certain implicitly included headers as well.

julia> c"""
         #include <mylib/header.h>
  • all C types are defined in Julia
  • C function and global variable bindings defined
  • the C API is documented and exported by the enclosing module

All done in just a few lines of code! Take a look at the complete example below or continue reading to learn about some more details.

Some gory details

The C expressions are parsed and immediately converted to Julia code. In fact, the generated Julia code can be inspected using @macroexpand, like this:

julia> @macroexpand c"""
         struct S;
         struct T {
           int i;
           struct S *s;
           struct T *t;
         extern void func(struct S *s, struct T t);

In order to support the fully automatic conversion and avoid name collisions, the names of C types or functions are mangled a bit to work in Julia. Therefore everything generated by CBinding.jl can be accessed with the c"..." string macro (more about this below) to indicate that it lives in C-land. As an example, the function func above is available in Julia as c"func". It is possible to store the generated bindings to more user-friendly names (this can sometimes be automated, see the j option). Placing each C declaration in its own macro helps when doing this manually, like:

julia> const S = c"""
         struct S;

julia> const T = c"""
         struct T {
           int i;
           struct S *s;
           struct T *t;

julia> c"""
         extern void func(struct S *s, struct T t);

Constructs from the standard C library headers are currently not being emitted by CBinding.jl, but other packages may be developed to provide a unified source for them. For now, dependencies on C library or other libraries should be placed before any C code blocks referencing them. Most often it is only a few using and const statements.

A complete example

Finally, a set of examples can be found at https://github.com/analytech-solutions/ExamplesUsingCBinding.jl, but here is a generalized example of what a package using CBinding.jl might look like:

module LibFoo
  module libfoo
    import Foo_jll
    using CBinding
    # libfoo has libbar as a dep, and LibBar has bindings for it
    using LibBar: libbar
    # set up the parser
      incdir = joinpath(Foo_jll.artifact_dir, "include")
      libdir = dirname(Foo_jll.libfoo_path)
      c`-std=c99 -fparse-all-comments -I$(incdir) -L$(libdir) -lfoo`
    # libfoo refers to some std C sized types (eventually made available with something like `using C99`)
    const c"int32_t"  = Int32
    const c"int64_t"  = Int64
    const c"uint32_t" = UInt32
    const c"uint64_t" = UInt64
    # generate bindings for libfoo
      #include <libfoo/header-1.h>
      #include <libfoo/header-2.h>
    # any other bindings not in headers
      struct FooStruct {
        struct BarStruct bs;
      extern struct FooStruct *foo_like_its_the_80s(int i);
  # high-level Julian interface to libfoo
  using CBinding
  using .libfoo
  function foo(i)
    ptr = c"foo_like_its_the_80s"(Cint(i-1))
      return JulianFoo(ptr[])

Options for c"..."

The string macro has some options to handle more complex use cases. Occasionally it is necessary to include or define C code that is just a dependency and should not be exported or perhaps excluded from the generated bindings altogether. These kinds of situations can be handled with combinations of the following string macro suffixes.

  • d - defer conversion of the C code block; successive blocks marked with d will keep deferring until a block without it (its options will be used for processing the deferred blocks)
  • f - don't create bindings for extern functions
  • i - also parse implicitly included headers that are related (in the same directory or subdirectories) to explicitly included headers
  • j - provide additional bindings using Julian names (name collisions likely)
  • J - provide additional bindings using Julian names with annotated user-defined types (using struct_, union_, or enum_ prefixes)
  • m - skip conversion of C macros
  • n - show warnings for macros or inline functions that are skipped (and other conversion issues)
  • p - mark the C code as "private" content that will not be exported
  • q - quietly parse the block of C code, suppressing any compiler/linker messages
  • r - the C code is only a reference to something in C-land and bindings are not to be generated
  • s - skip processing of this block of C code
  • t - skip conversion of C types
  • u - leave this block of C code undocumented
  • v - don't create bindings for extern variables
  • w - create bindings for inline functions by using wrapper libraries (somewhat experimental)
julia> c"""
         #include <stdio.h>  // provides FILE type, but skip emitting bindings for this block

julia> c"""
         struct File {  // do not include this type in module exports, and suppress compiler messages
           FILE *f;

Using CBinding.jl-generated bindings

The c"..." string macro can be used to refer to any of the types, global variables, or functions generated by CBinding.jl. When simply referencing the C content, setting up a compiler context (i.e. using c`...`) is not necessary.

The c"..." string macro can take on two meanings depending on the content placed in it. So to guarantee it is interpreted as a reference to something in C, rather than a block of C code to create bindings with, include an r in the string macro options.

julia> module MyLib  # generally some C bindings are defined elsewhere
         using CBinding
         c`-std=c99 -Wall -Imy/include`
           struct S;
           struct T {
             int i;
             struct S *s;
             struct T *t;
           extern void func(struct S *s, struct T t);

julia> using CBinding, .MyLib

julia> c"struct T" <: Cstruct

julia> c"struct T"r <: Cstruct  # use 'r' option to guarantee it is treated as a reference

julia> t = c"struct T"(i = 123);

julia> t.i

The user-defined types (enum, struct, and union) are referenced just as they are in C (e.g. c"enum E", c"struct S", and c"union U"). All other types, pointers, arrays, global variables, enumeration constants, functions, etc. are also referenced just as they are in C. Here is a quick reference for C string macro usage:

  • c"int" - the Cint type
  • c"int[2]" - a length-2 static array of Cint's
  • c"int[2][4]" - a length-2 static array of length-4 static arrays of Cint's
  • c"int *" - pointer to a Cint
  • c"int **" - pointer to a pointer to a Cint
  • c"int const **" - pointer to a pointer to a read-only Cint
  • c"enum MyUnion" - a user-defined C enum type
  • c"union MyUnion" - a user-defined C union type
  • c"struct MyStruct" - a user-defined C struct type
  • c"struct MyStruct *" - a pointer to a user-defined C struct type
  • c"struct MyStruct[2]" - a length-2 static array of user-defined C struct type
  • c"MyStruct" - a user-defined typedef-ed type
  • c"MyStruct *" - a pointer to a user-defined typedef-ed type
  • c"printf" - a C function (specifically the printf function)
  • c"int (*)(int, int)" - a function pointer
  • c"int (*)(char const *, ...)" - a variadic function pointer

The following examples demonstrate how to refer to C-land content that resides in other modules and is not exported/imported:

  • c"SomeModule.SubModule.enum MyUnion"
  • c"SomeModule.SubModule.struct MyStruct *"
  • c"SomeModule.SubModule.printf"
  • c"int (*)(Some.Other.Module.struct MyStruct *, ...)"

The C string macro can also be used to expose Julia content to C-land.

julia> const c"IntPtr" = Cptr{Cint};

julia> c"void (*)(IntPtr, IntPtr *, IntPtr[2])" <: Cptr{<:Cfunction}

Type qualifiers are carried over from the C code. As an example, int const * is a pointer to a read-only integer in is represented by CBinding.jl as the type Cptr{Cconst{Cint}}. The unqualifiedtype(T) can be used to strip away the type qualifiers to get to the core type, so unqualifiedtype(Cconst{Cint}) === Cint.

As detailed below, the bitstype(T) function can be used to acquire the concrete bits type of user-defined C types as well.

Working with aggregate types and sized arrays

User-defined aggregate types (struct and union) have several ways to be constructed:

  • t = c"struct T"() - zero-ed immutable object
  • t = c"struct T"(i = 123) - zero-ed immutable object with field i initialized to 123
  • t = c"struct T"(t, i = 321) - copy of t with field i initialized to 321

These objects are immutable and changing fields will have no effect, so a copy must be constructed with the desired field overrides or pointers must be used. Nested field access is transparent, and performance should match that of accessing fields within standard Julia immutable structs.

Statically-sized arrays (i.e. c"typedef int IntArray[4];") can be constructed:

  • t = c"IntArray"() - zero-ed immutable array
  • t = c"IntArray"(1, 2) - zero-ed immutable array with first 2 elements initialized to 1 and 2
  • t = c"IntArray"(t, 3) - copy of t with first element initialized to 3
  • t = c"IntArray"(t, 4 => 123) - copy of t with 4th element initialized to 123

Constructors for both aggregates and arrays can also accept nested Tuple and NamedTuple arguments which get splatted appropriately into the respective field's constructor. A comprehensive example of constructing a complex C type and accessing fields/elements is shown below:

julia> c`` ; c"""
         struct A {
           struct {
             int i;
           struct {
             struct {
               int i;
             } c[2];
           } b;

julia> a = c"struct A"();

julia> a.i

julia> a.b.c[2].i

julia> a = c"struct A"(i = 123, b = (c = ((i = 321,), (i = 654,)),));

julia> a.i

julia> a.b.c[2].i

Working with pointers

CBinding.jl also works elegantly with pointers to aggregate types. Pointers are followed through fields and array elements as they are accessed, and they can be dereferenced with ptr[] or written to with ptr[] = val.

julia> ptr = Libc.malloc(a);  # allocate a `struct A` as a copy of `a`

julia> ptr.i

julia> ptr.i[]

julia> ptr.b.c[2].i

julia> ptr.b.c[2].i[]

julia> ptr.b.c[2].i[] = 42

julia> Libc.free(ptr)  # deallocate it

An exception to the rule is bitfields. It is not possible to refer to bitfields with a pointer, so access to bitfields is automatically dereferenced.

Using global variable and function bindings

Bindings to global variables also behave as if they are pointers, and must be dereferenced to be read or written. Fields and elements can be followed through the same as with pointers. Bindings to functions can be called directly, and getting the pointer to a one can be done with the "dereferencing" (func[]) syntax in case a bound function must be used as a callback function.

julia> c"func"(Cint(1), Cint(2));  # call the C function directly

julia> funcptr = c"func"[]
Cptr{Cfunction{Int32, Tuple{Int32, Int32}, :cdecl}}(0x00007f8f50722b10)

julia> funcptr(Cint(1), Cint(2));  # call the C function pointer

Using Julia functions in C

Providing a Julia method to C as a callback function has never been easier! Just pass it as an argument to the CBinding.jl function binding or function pointer. Assuming a binding to a C function, like void set_callback(int (*cb)(int, int)) exists:

julia> function myadd(a, b)  # a callback function to give to C
         return a+b

julia> c"set_callback"(myadd)  # that's it!

julia> function saferadd(a::Cint, b::Cint)::Cint  # a safer callback function might require type paranoia
         return a+b

julia> c"set_callback"(saferadd)

Getting help

Unless explicitly disabled, the generated bindings include doc-strings. An attempt is made at converting any structured comments from the C blocks into somewhat equivalent doc-strings, as this example illustrates:

help?> libsdl2.SDL_CreateRGBSurface
  extern SDL_Surface *SDL_CreateRGBSurface(Uint32 flags, int width, int height, int depth, Uint32 Rmask, Uint32 Gmask, Uint32 Bmask, Uint32 Amask)

  Defined at SDL_surface.h:130 (file:///usr/include/SDL2/SDL_surface.h)

   Allocate and free an RGB surface.


   If the depth is 4 or 8 bits, an empty palette is allocated for the surface. If the depth is greater than 8 bits, the pixel format is set
  using the flags '[RGB]mask'.

   If the function runs out of memory, it will return NULL.

  ============flags: The flags are obsolete and should be set to 0.width: The width in pixels of the surface to create.height: The height in pixels of the surface to create.depth: The depth in bits of the surface to create.Rmask: The red mask of the surface to create.Gmask: The green mask of the surface to create.Bmask: The blue mask of the surface to create.Amask: The alpha mask of the surface to create.

However, if such exquisite documentation cannot be generated, the doc-string simply conveys the item's original C definition:

help?> libclang.clang_visitChildren
  unsigned int clang_visitChildren(CXCursor parent, CXCursorVisitor visitor, CXClientData client_data)

  Defined at Index.h:4189 (file:///usr/include/clang-c/Index.h)

help?> libclang.CXCursor
  struct {
      enum CXCursorKind kind;
      int xdata;
      const void *data[3];

  Defined at Index.h:2664 (file:///usr/include/clang-c/Index.h)

Check for comments near the referenced definition location for C documentation that libclang failed to associate with the binding.

Any gotchas?

Since Julia does not yet provide incomplete type (please voice your support of the feature here: https://github.com/JuliaLang/julia/issues/269), abstract types are used to allow forward declarations in C. Therefore, referencing C types usually refers to the abstract type which can have significant implications when creating Julia arrays, using ccall, etc. The following example illustrates this kind of unexpected behavior:

julia> struct X

julia> const Y = c"""
         struct Y {
           int i;

julia> [X(123)] isa Vector{X}

julia> [Y(i=123)] isa Vector{Y}

julia> [Y(i=123)] isa Vector{bitstype(Y)}

The bitstype(T) function can be used to acquire the concrete bits type of any C type when the distinction matters.

Another implementation detail worth noting is that function bindings are brought into Julia as singleton constants, not as actual functions. This approach allows a user to obtain function pointers from C functions in case one must be used as a callback function. Therefore, attaching other methods to a bound C function is not possible.

It is also sometimes necessary to use the c"..." mangled names directly in Julia (for instance in the REPL help mode). Until consistent, universal support for the string macro is available, the mangled names can be used directly as var"c\"...\"", like help?> var"c\"struct Y\"".

When a C function has a statically-sized array as an argument in its signature, the semantics of C is to treat the argument as a pointer instead. Therefore, the statically-sized array in the binding signature will be lowered to a pointer for the underlying ccall. A user can then pass any argument (array, pointer, etc.) that is compatible with a pointer argument type (not the anticipated statically-sized array type).

Some helpful tips

Thorough tutorials and examples should be developed illustrating the countless scenarios that could be encountered when interfacing C from Julia in an automated manner. However, until such a body of a work is available, the following list of hopefully helpful comments will have to suffice:

  • Documentation comments are not converted? Try adding -fparse-all-comments to your compiler context command.
  • Encounter syntax/types that are not yet supported errors? Try disabling compiler extensions (e.g. -fno-blocks) or specifying a language standard (such as -std=c99).
  • Are some functions or macros missing from the generated bindings? See warnings if CBinding.jl skips anything by using the 'n' string macro option (like c"..."n).
  • Are there undefined references to C library items (such as uint8_t, FILE, or va_list)? Until C library modules are published, define these symbols using existing Julia types: const c"uint8_t" = UInt8, const c"FILE" = Cvoid, const c"va_list" = Cvoid, etc.