Generator Tutorial

Tutorial on wrapping a JLL package

In most situations, Clang.jl is used to export a Julia interface to a C library managed by a JLL package. A JLL package wraps an artifact which provides a shared library that can be called with the ccall syntax and headers suitable for a C compiler. Clang.jl can translate the C headers into Julia files that can be directly used like normal Julia functions and types.

The general workflow of wrapping a JLL package is as follows.

  1. Locate the C headers relative to the artifact directory.
  2. Find the compiler flags needed to parse these headers.
  3. Create a .toml file with generator options.
  4. Build a context with the above three and run.
  5. Test and troubleshoot the wrapper.

Create a default generator

A generator context consists of a list of headers, a list of compiler flags, and generator options. The example below creates a typical context and runs the generator.

using Clang.Generators
using Clang.LibClang.Clang_jll


include_dir = normpath(Clang_jll.artifact_dir, "include")

# wrapper generator options
options = load_options(joinpath(@__DIR__, "generator.toml"))

# add compiler flags, e.g. "-DXXXXXXXXX"
args = get_default_args()
push!(args, "-I$include_dir")

# only wrap libclang headers in include/clang-c
header_dir = joinpath(include_dir, "clang-c")
headers = [joinpath(header_dir, header) for header in readdir(header_dir) if endswith(header, ".h")]

# create context
ctx = create_context(headers, args, options)

# run generator

You can also use the experimental detect_headers function to automatically detect top-level headers in the directory.

headers = detect_headers(header_dir, args)

You also need an options file generator.toml that to make this script work, you can refer to this toml file for a reference.

Skipping specific symbols

The C header may contain some symbols that are not correctly handled by Clang.jl or may need manual wrapping. For example, julia provides tm as Libc.TmStruct, so you may not want to map it to a new struct. As a workaround, you can skip these symbols. After that, if this symbol is needed, you can add it back in the prologue. Prologue is specified by the prologue_file_path option.

  • Add the symbol to output_ignorelist to avoid it from being wrapped.
  • If the symbol is in system headers and causes Clang.jl to error before printing, apart from posting an issue, write @add_def symbol_name before generating to suppress it from being wrapped.

Rewrite expressions before printing

You can also modify the generated wrapped before it is printed. Clang.jl separates the building process into generating and printing processes. You can run these two processes separately and rewrite the expressions before printing.

# build without printing so we can do custom rewriting

# custom rewriter
function rewrite!(e::Expr)

function rewrite!(dag::ExprDAG)
    for node in get_nodes(dag)
        for expr in get_exprs(node)


# print

Multi-platform configuration

Some headers may contain system-dependent symbols such as long or char, or system-independent symbols may be resolved to system-dependent ones. For example, time_t is usually just a 64-bit unsigned integer, but implementations may conditionally implement it as long or long long, which is not portable. You can skip these symbols and add them back manually as in Skipping specific symbols. If the differences are too large to be manually fixed, you can generate wrappers for each platform as in LibClang.jl.

Variadic Function

With the help of @ccall macro, variadic C functions can be called from Julia. For example, @ccall printf("%d\n"::Cstring; 123::Cint)::Cint can be used to call the C function printf. Note that those arguments after the semicolon ; are variadic arguments.

If wrap_variadic_function in codegen section of options is set to true, Clang.jl will generate wrappers for variadic C functions. For example, printf will be wrapped as follows.

@generated function printf(fmt, va_list...)
        :(@ccall(libexample.printf(fmt::Ptr{Cchar}; $(to_c_type_pairs(va_list)...))::Cint))

It can be called just like normal Julia functions without specifying types: LibExample.printf("%d\n", 123).


Although variadic functions are supported, the C type va_list cannot be used from Julia.

Type Correspondence

However, variadic C functions must be called with the correct argument types. The most useful ones are listed below.

C typeccall signatureJulia type
Integers and floating point numbersthe same typethe same type
Struct Ta concrete Julia struct T with the same layoutT
Pointer (T*)Ref{T} or Ptr{T}Ref{T} or Ptr{T} or any array type
String (char*)Cstring or Ptr{Cchar}String

Ref is not a concrete type but an abstract type in Julia. For example, Ref(1) is Base.RefValue(1), which cannot be directly passed to C.

As observed from the table, if you want to pass strings or arrays to C, you need to annotate the type as Ptr{T} or Ref{T} (or Cstring). Otherwise, the struct that represents the String or Array type instead of the buffer itself will be passed. There are two methods to pass arguments of these types:

  • Directly use the @ccall macro: @ccall printf("%s\n"; "hello"::Cstring)::Cint. You can also create wrappers for common use cases of this.
  • Overload to_c_type to map Julia type to correct ccall signature type: add to_c_type(::Type{String}) = Cstring to prologue (prologue can be added by setting prologue_file_path in options). Then all arguments of type String will be annotated as Cstring.

The above type correspondence can be implemented by including the following lines in the prologue.

to_c_type(::Type{<:AbstractString}) = Cstring # or Ptr{Cchar}
to_c_type(t::Type{<:Union{AbstractArray,Ref}}) = Ptr{eltype(t)}

For a complete tutorial on calling C functions, refer to Calling C and Fortran Code in the Julia manual.