Movatterモバイル変換


[0]ホーム

URL:


GitHub

Calling C and Fortran Code

Though most code can be written in Julia, there are many high-quality, mature libraries for numerical computing already written in C and Fortran. To allow easy use of this existing code, Julia makes it simple and efficient to call C and Fortran functions. Julia has a "no boilerplate" philosophy: functions can be called directly from Julia without any "glue" code, code generation, or compilation – even from the interactive prompt. This is accomplished just by making an appropriate call with the@ccall macro (or the less convenientccall syntax, see theccall syntax section).

The code to be called must be available as a shared library. Most C and Fortran libraries ship compiled as shared libraries already, but if you are compiling the code yourself using GCC (or Clang), you will need to use the-shared and-fPIC options. The machine instructions generated by Julia's JIT are the same as a native C call would be, so the resulting overhead is the same as calling a library function from C code.[1]

By default, Fortran compilersgenerate mangled names (for example, converting function names to lowercase or uppercase, often appending an underscore), and so to call a Fortran function you must pass the mangled identifier corresponding to the rule followed by your Fortran compiler. Also, when calling a Fortran function, all inputs must be passed as pointers to allocated values on the heap or stack. This applies not only to arrays and other mutable objects which are normally heap-allocated, but also to scalar values such as integers and floats which are normally stack-allocated and commonly passed in registers when using C or Julia calling conventions.

The syntax for@ccall to generate a call to the library function is:

  @ccall library.function_name(argvalue1::argtype1, ...)::returntype  @ccall function_name(argvalue1::argtype1, ...)::returntype  @ccall $function_pointer(argvalue1::argtype1, ...)::returntype

wherelibrary is a string constant or literal (but seeNon-constant Function Specifications below). The library may be omitted, in which case the function name is resolved in the current process. This form can be used to call C library functions, functions in the Julia runtime, or functions in an application linked to Julia. The full path to the library may also be specified. Alternatively,@ccall may also be used to call a function pointer$function_pointer, such as one returned byLibdl.dlsym. Theargtypes corresponds to the C-function signature and theargvalues are the actual argument values to be passed to the function.

Note

See below for how tomap C types to Julia types.

As a complete but simple example, the following calls theclock function from the standard C library on most Unix-derived systems:

julia> t = @ccall clock()::Int322292761julia> typeof(t)Int32

clock takes no arguments and returns anInt32. To call thegetenv function to get a pointer to the value of an environment variable, one makes a call like this:

julia> path = @ccall getenv("SHELL"::Cstring)::CstringCstring(@0x00007fff5fbffc45)julia> unsafe_string(path)"/bin/bash"

In practice, especially when providing reusable functionality, one generally wraps@ccall uses in Julia functions that set up arguments and then check for errors in whatever manner the C or Fortran function specifies. And if an error occurs it is thrown as a normal Julia exception. This is especially important since C and Fortran APIs are notoriously inconsistent about how they indicate error conditions. For example, thegetenv C library function is wrapped in the following Julia function, which is a simplified version of the actual definition fromenv.jl:

function getenv(var::AbstractString)    val = @ccall getenv(var::Cstring)::Cstring    if val == C_NULL        error("getenv: undefined variable: ", var)    end    return unsafe_string(val)end

The Cgetenv function indicates an error by returningC_NULL, but other standard C functions indicate errors in different ways, including by returning -1, 0, 1, and other special values. This wrapper throws an exception indicating the problem if the caller tries to get a non-existent environment variable:

julia> getenv("SHELL")"/bin/bash"julia> getenv("FOOBAR")ERROR: getenv: undefined variable: FOOBAR

Here is a slightly more complex example that discovers the local machine's hostname.

function gethostname()    hostname = Vector{UInt8}(undef, 256) # MAXHOSTNAMELEN    err = @ccall gethostname(hostname::Ptr{UInt8}, sizeof(hostname)::Csize_t)::Int32    Base.systemerror("gethostname", err != 0)    hostname[end] = 0 # ensure null-termination    return GC.@preserve hostname unsafe_string(pointer(hostname))end

This example first allocates an array of bytes. It then calls the C library functiongethostname to populate the array with the hostname. Finally, it takes a pointer to the hostname buffer, and converts the pointer to a Julia string, assuming that it is a null-terminated C string.

It is common for C libraries to use this pattern of requiring the caller to allocate memory to be passed to the callee and populated. Allocation of memory from Julia like this is generally accomplished by creating an uninitialized array and passing a pointer to its data to the C function. This is why we don't use theCstring type here: as the array is uninitialized, it could contain null bytes. Converting to aCstring as part of the@ccall checks for contained null bytes and could therefore throw a conversion error.

Dereferencingpointer(hostname) withunsafe_string is an unsafe operation as it requires access to the memory allocated forhostname that may have been in the meanwhile garbage collected. The macroGC.@preserve prevents this from happening and therefore accessing an invalid memory location.

Finally, here is an example of specifying a library via a path. We create a shared library with the following content

#include <stdio.h>void say_y(int y){    printf("Hello from C: got y = %d.\n", y);}

and compile it withgcc -fPIC -shared -o mylib.so mylib.c. It can then be called by specifying the (absolute) path as the library name:

julia> @ccall "./mylib.so".say_y(5::Cint)::CvoidHello from C: got y = 5.

Creating C-Compatible Julia Function Pointers

It is possible to pass Julia functions to native C functions that accept function pointer arguments. For example, to match C prototypes of the form:

typedef returntype (*functiontype)(argumenttype, ...)

The macro@cfunction generates the C-compatible function pointer for a call to a Julia function. The arguments to@cfunction are:

  1. A Julia function
  2. The function's return type
  3. A tuple of input types, corresponding to the function signature
Note

As with@ccall, the return type and the input types must be literal constants.

Note

Currently, only the platform-default C calling convention is supported. This means that@cfunction-generated pointers cannot be used in calls where WINAPI expects astdcall function on 32-bit Windows, but can be used on WIN64 (wherestdcall is unified with the C calling convention).

Note

Callback functions exposed via@cfunction should not throw errors, as that will return control to the Julia runtime unexpectedly and may leave the program in an undefined state.

A classic example is the standard C libraryqsort function, declared as:

void qsort(void *base, size_t nitems, size_t size,           int (*compare)(const void*, const void*));

Thebase argument is a pointer to an array of lengthnitems, with elements ofsize bytes each.compare is a callback function which takes pointers to two elementsa andb and returns an integer less/greater than zero ifa should appear before/afterb (or zero if any order is permitted).

Now, suppose that we have a 1-d arrayA of values in Julia that we want to sort using theqsort function (rather than Julia's built-insort function). Before we consider callingqsort and passing arguments, we need to write a comparison function:

julia> function mycompare(a, b)::Cint           return (a < b) ? -1 : ((a > b) ? +1 : 0)       end;

qsort expects a comparison function that return a Cint, so we annotate the return type to beCint.

In order to pass this function to C, we obtain its address using the macro@cfunction:

julia> mycompare_c = @cfunction(mycompare, Cint, (Ref{Cdouble}, Ref{Cdouble}));

@cfunction requires three arguments: the Julia function (mycompare), the return type (Cint), and a literal tuple of the input argument types, in this case to sort an array ofCdouble (Float64) elements.

The final call toqsort looks like this:

julia> A = [1.3, -2.7, 4.4, 3.1];julia> @ccall qsort(A::Ptr{Cdouble}, length(A)::Csize_t, sizeof(eltype(A))::Csize_t, mycompare_c::Ptr{Cvoid})::Cvoidjulia> A4-element Vector{Float64}: -2.7  1.3  3.1  4.4

As the example shows, the original Julia arrayA has now been sorted:[-2.7, 1.3, 3.1, 4.4]. Note that Juliatakes care of converting the array to aPtr{Cdouble}), computing the size of the element type in bytes, and so on.

For fun, try inserting aprintln("mycompare($a, $b)") line intomycompare, which will allow you to see the comparisons thatqsort is performing (and to verify that it is really calling the Julia function that you passed to it).

Mapping C Types to Julia

It is critical to exactly match the declared C type with its declaration in Julia. Inconsistencies can cause code that works correctly on one system to fail or produce indeterminate results on a different system.

Note that no C header files are used anywhere in the process of calling C functions: you are responsible for making sure that your Julia types and call signatures accurately reflect those in the C header file.[2]

Automatic Type Conversion

Julia automatically inserts calls to theBase.cconvert function to convert each argument to the specified type. For example, the following call:

@ccall "libfoo".foo(x::Int32, y::Float64)::Cvoid

will behave as if it were written like this:

c_x = Base.cconvert(Int32, x)c_y = Base.cconvert(Float64, y)GC.@preserve c_x c_y begin    @ccall "libfoo".foo(        Base.unsafe_convert(Int32, c_x)::Int32,        Base.unsafe_convert(Float64, c_y)::Float64    )::Cvoidend

Base.cconvert normally just callsconvert, but can be defined to return an arbitrary new object more appropriate for passing to C. This should be used to perform all allocations of memory that will be accessed by the C code. For example, this is used to convert anArray of objects (e.g. strings) to an array of pointers.

Base.unsafe_convert handles conversion toPtr types. It is considered unsafe because converting an object to a native pointer can hide the object from the garbage collector, causing it to be freed prematurely.

Type Correspondences

First, let's review some relevant Julia type terminology:

Syntax / KeywordExampleDescription
mutable structBitSet"Leaf Type" :: A group of related data that includes a type-tag, is managed by the Julia GC, and is defined by object-identity. The type parameters of a leaf type must be fully defined (noTypeVars are allowed) in order for the instance to be constructed.
abstract typeAny,AbstractArray{T, N},Complex{T}"Super Type" :: A super-type (not a leaf-type) that cannot be instantiated, but can be used to describe a group of types.
T{A}Vector{Int}"Type Parameter" :: A specialization of a type (typically used for dispatch or storage optimization).
"TypeVar" :: TheT in the type parameter declaration is referred to as a TypeVar (short for type variable).
primitive typeInt,Float64"Primitive Type" :: A type with no fields, but a size. It is stored and defined by-value.
structPair{Int, Int}"Struct" :: A type with all fields defined to be constant. It is defined by-value, and may be stored with a type-tag.
ComplexF64 (isbits)"Is-Bits" :: Aprimitive type, or astruct type where all fields are otherisbits types. It is defined by-value, and is stored without a type-tag.
struct ...; endnothing"Singleton" :: a Leaf Type or Struct with no fields.
(...) ortuple(...)(1, 2, 3)"Tuple" :: an immutable data-structure similar to an anonymous struct type, or a constant array. Represented as either an array or a struct.

Bits Types

There are several special types to be aware of, as no other type can be defined to behave the same:

On all systems we currently support, basic C/C++ value types may be translated to Julia types as follows. Every C type also has a corresponding Julia type with the same name, prefixed by C. This can help when writing portable code (and remembering that anint in C is not the same as anInt in Julia).

System Independent Types

C nameFortran nameStandard Julia AliasJulia Base Type
unsigned charCHARACTERCucharUInt8
bool (_Bool in C99+)CucharUInt8
shortINTEGER*2,LOGICAL*2CshortInt16
unsigned shortCushortUInt16
int,BOOL (C, typical)INTEGER*4,LOGICAL*4CintInt32
unsigned intCuintUInt32
long longINTEGER*8,LOGICAL*8ClonglongInt64
unsigned long longCulonglongUInt64
intmax_tCintmax_tInt64
uintmax_tCuintmax_tUInt64
floatREAL*4iCfloatFloat32
doubleREAL*8CdoubleFloat64
complex floatCOMPLEX*8ComplexF32Complex{Float32}
complex doubleCOMPLEX*16ComplexF64Complex{Float64}
ptrdiff_tCptrdiff_tInt
ssize_tCssize_tInt
size_tCsize_tUInt
voidCvoid
void and[[noreturn]] or_NoreturnUnion{}
void*Ptr{Cvoid} (or similarlyRef{Cvoid})
T* (where T represents an appropriately defined type)Ref{T} (T may be safely mutated only if T is an isbits type)
char* (orchar[], e.g. a string)CHARACTER*NCstring if null-terminated, orPtr{UInt8} if not
char** (or*char[])Ptr{Ptr{UInt8}}
jl_value_t* (any Julia Type)Any
jl_value_t* const* (a reference to a Julia value)Ref{Any} (const, since mutation would require a write barrier, which is not possible to insert correctly)
va_argNot supported
... (variadic function specification)T... (whereT is one of the above types, when using theccall function)
... (variadic function specification); va_arg1::T, va_arg2::S, etc. (only supported with@ccall macro)

TheCstring type is essentially a synonym forPtr{UInt8}, except the conversion toCstring throws an error if the Julia string contains any embedded null characters (which would cause the string to be silently truncated if the C routine treats null as the terminator). If you are passing achar* to a C routine that does not assume null termination (e.g. because you pass an explicit string length), or if you know for certain that your Julia string does not contain null and want to skip the check, you can usePtr{UInt8} as the argument type.Cstring can also be used as theccall return type, but in that case it obviously does not introduce any extra checks and is only meant to improve the readability of the call.

System Dependent Types

C nameStandard Julia AliasJulia Base Type
charCcharInt8 (x86, x86_64),UInt8 (powerpc, arm)
longClongInt (UNIX),Int32 (Windows)
unsigned longCulongUInt (UNIX),UInt32 (Windows)
wchar_tCwchar_tInt32 (UNIX),UInt16 (Windows)
Note

When calling Fortran, all inputs must be passed by pointers to heap- or stack-allocated values, so all type correspondences above should contain an additionalPtr{..} orRef{..} wrapper around their type specification.

Warning

For string arguments (char*) the Julia type should beCstring (if null-terminated data is expected), or eitherPtr{Cchar} orPtr{UInt8} otherwise (these two pointer types have the same effect), as described above, notString. Similarly, for array arguments (T[] orT*), the Julia type should again bePtr{T}, notVector{T}.

Warning

Julia'sChar type is 32 bits, which is not the same as the wide-character type (wchar_t orwint_t) on all platforms.

Warning

A return type ofUnion{} means the function will not return, i.e., C++11[[noreturn]] or C11_Noreturn (e.g.jl_throw orlongjmp). Do not use this for functions that return no value (void) but do return, for those, useCvoid instead.

Note

Forwchar_t* arguments, the Julia type should beCwstring (if the C routine expects a null-terminated string), orPtr{Cwchar_t} otherwise. Note also that UTF-8 string data in Julia is internally null-terminated, so it can be passed to C functions expecting null-terminated data without making a copy (but using theCwstring type will cause an error to be thrown if the string itself contains null characters).

Note

C functions that take an argument of typechar** can be called by using aPtr{Ptr{UInt8}} type within Julia. For example, C functions of the form:

int main(int argc, char **argv);

can be called via the following Julia code:

argv = [ "a.out", "arg1", "arg2" ]@ccall main(length(argv)::Int32, argv::Ptr{Ptr{UInt8}})::Int32
Note

For Fortran functions taking variable length strings of typecharacter(len=*) the string lengths are provided ashidden arguments. Type and position of these arguments in the list are compiler specific, where compiler vendors usually default to usingCsize_t as type and append the hidden arguments at the end of the argument list. While this behaviour is fixed for some compilers (GNU), othersoptionally permit placing hidden arguments directly after the character argument (Intel, PGI). For example, Fortran subroutines of the form

subroutine test(str1, str2)character(len=*) :: str1,str2

can be called via the following Julia code, where the lengths are appended

str1 = "foo"str2 = "bar"ccall(:test, Cvoid, (Ptr{UInt8}, Ptr{UInt8}, Csize_t, Csize_t),                    str1, str2, sizeof(str1), sizeof(str2))
Warning

Fortran compilersmay also add other hidden arguments for pointers, assumed-shape (:) and assumed-size (*) arrays. Such behaviour can be avoided by usingISO_C_BINDING and includingbind(c) in the definition of the subroutine, which is strongly recommended for interoperable code. In this case, there will be no hidden arguments, at the cost of some language features (e.g. onlycharacter(len=1) will be permitted to pass strings).

Note

A C function declared to returnCvoid will return the valuenothing in Julia.

Struct Type Correspondences

Composite types such asstruct in C orTYPE in Fortran90 (orSTRUCTURE /RECORD in some variants of F77), can be mirrored in Julia by creating astruct definition with the same field layout.

When used recursively,isbits types are stored inline. All other types are stored as a pointer to the data. When mirroring a struct used by-value inside another struct in C, it is imperative that you do not attempt to manually copy the fields over, as this will not preserve the correct field alignment. Instead, declare anisbits struct type and use that instead. Unnamed structs are not possible in the translation to Julia.

Packed structs and union declarations are not supported by Julia.

You can get an approximation of aunion if you know, a priori, the field that will have the greatest size (potentially including padding). When translating your fields to Julia, declare the Julia field to be only of that type.

Arrays of parameters can be expressed withNTuple. For example, the struct in C notation is written as

struct B {    int A[3];};b_a_2 = B.A[2];

can be written in Julia as

struct B    A::NTuple{3, Cint}endb_a_2 = B.A[3]  # note the difference in indexing (1-based in Julia, 0-based in C)

Arrays of unknown size (C99-compliant variable length structs specified by[] or[0]) are not directly supported. Often the best way to deal with these is to deal with the byte offsets directly. For example, if a C library declared a proper string type and returned a pointer to it:

struct String {    int strlen;    char data[];};

In Julia, we can access the parts independently to make a copy of that string:

str = from_c::Ptr{Cvoid}len = unsafe_load(Ptr{Cint}(str))unsafe_string(str + Core.sizeof(Cint), len)

Type Parameters

The type arguments to@ccall and@cfunction are evaluated statically, when the method containing the usage is defined. They therefore must take the form of a literal tuple, not a variable, and cannot reference local variables.

This may sound like a strange restriction, but remember that since C is not a dynamic language like Julia, its functions can only accept argument types with a statically-known, fixed signature.

However, while the type layout must be known statically to compute the intended C ABI, the static parameters of the function are considered to be part of this static environment. The static parameters of the function may be used as type parameters in the call signature, as long as they don't affect the layout of the type. For example,f(x::T) where {T} = @ccall valid(x::Ptr{T})::Ptr{T} is valid, sincePtr is always a word-size primitive type. But,g(x::T) where {T} = @ccall notvalid(x::T)::T is not valid, since the type layout ofT is not known statically.

SIMD Values

If a C/C++ routine has an argument or return value that is a native SIMD type, the corresponding Julia type is a homogeneous tuple ofVecElement that naturally maps to the SIMD type. Specifically:

  • The tuple must be the same size and elements as the SIMD type. For example, a tuple representing an__m128 on x86 must have a size of 16 bytes and Float32 elements.
  • The element type of the tuple must be an instance ofVecElement{T} whereT is a primitive type with a power-of-two number of bytes (e.g. 1, 2, 4, 8, 16, etc) such as Int8 or Float64.

For instance, consider this C routine that uses AVX intrinsics:

#include <immintrin.h>__m256 dist( __m256 a, __m256 b ) {    return _mm256_sqrt_ps(_mm256_add_ps(_mm256_mul_ps(a, a),                                        _mm256_mul_ps(b, b)));}

The following Julia code callsdist usingccall:

const m256 = NTuple{8, VecElement{Float32}}a = m256(ntuple(i -> VecElement(sin(Float32(i))), 8))b = m256(ntuple(i -> VecElement(cos(Float32(i))), 8))function call_dist(a::m256, b::m256)    @ccall "libdist".dist(a::m256, b::m256)::m256endprintln(call_dist(a,b))

The host machine must have the requisite SIMD registers. For example, the code above will not work on hosts without AVX support.

Memory Ownership

malloc/free

Memory allocation and deallocation of such objects must be handled by calls to the appropriate cleanup routines in the libraries being used, just like in any C program. Do not try to free an object received from a C library withLibc.free in Julia, as this may result in thefree function being called via the wrong library and cause the process to abort. The reverse (passing an object allocated in Julia to be freed by an external library) is equally invalid.

When to useT,Ptr{T} andRef{T}

In Julia code wrapping calls to external C routines, ordinary (non-pointer) data should be declared to be of typeT inside the@ccall, as they are passed by value. For C code accepting pointers,Ref{T} should generally be used for the types of input arguments, allowing the use of pointers to memory managed by either Julia or C through the implicit call toBase.cconvert. In contrast, pointers returned by the C function called should be declared to be of the output typePtr{T}, reflecting that the memory pointed to is managed by C only. Pointers contained in C structs should be represented as fields of typePtr{T} within the corresponding Julia struct types designed to mimic the internal structure of corresponding C structs.

In Julia code wrapping calls to external Fortran routines, all input arguments should be declared as of typeRef{T}, as Fortran passes all variables by pointers to memory locations. The return type should either beCvoid for Fortran subroutines, or aT for Fortran functions returning the typeT.

Mapping C Functions to Julia

@ccall /@cfunction argument translation guide

For translating a C argument list to Julia:

@ccall /@cfunction return type translation guide

For translating a C return type to Julia:

Passing Pointers for Modifying Inputs

Because C doesn't support multiple return values, often C functions will take pointers to data that the function will modify. To accomplish this within a@ccall, you need to first encapsulate the value inside aRef{T} of the appropriate type. When you pass thisRef object as an argument, Julia will automatically pass a C pointer to the encapsulated data:

width = Ref{Cint}(0)range = Ref{Cfloat}(0)@ccall foo(width::Ref{Cint}, range::Ref{Cfloat})::Cvoid

Upon return, the contents ofwidth andrange can be retrieved (if they were changed byfoo) bywidth[] andrange[]; that is, they act like zero-dimensional arrays.

C Wrapper Examples

Let's start with a simple example of a C wrapper that returns aPtr type:

mutable struct gsl_permutationend# The corresponding C signature is#     gsl_permutation * gsl_permutation_alloc (size_t n);function permutation_alloc(n::Integer)    output_ptr = @ccall "libgsl".gsl_permutation_alloc(n::Csize_t)::Ptr{gsl_permutation}    if output_ptr == C_NULL # Could not allocate memory        throw(OutOfMemoryError())    end    return output_ptrend

TheGNU Scientific Library (here assumed to be accessible through:libgsl) defines an opaque pointer,gsl_permutation *, as the return type of the C functiongsl_permutation_alloc. As user code never has to look inside thegsl_permutation struct, the corresponding Julia wrapper simply needs a new type declaration,gsl_permutation, that has no internal fields and whose sole purpose is to be placed in the type parameter of aPtr type. The return type of theccall is declared asPtr{gsl_permutation}, since the memory allocated and pointed to byoutput_ptr is controlled by C.

The inputn is passed by value, and so the function's input signature is simply declared as::Csize_t without anyRef orPtr necessary. (If the wrapper was calling a Fortran function instead, the corresponding function input signature would instead be::Ref{Csize_t}, since Fortran variables are passed by pointers.) Furthermore,n can be any type that is convertible to aCsize_t integer; theccall implicitly callsBase.cconvert(Csize_t, n).

Here is a second example wrapping the corresponding destructor:

# The corresponding C signature is#     void gsl_permutation_free (gsl_permutation * p);function permutation_free(p::Ptr{gsl_permutation})    @ccall "libgsl".gsl_permutation_free(p::Ptr{gsl_permutation})::Cvoidend

Here is a third example passing Julia arrays:

# The corresponding C signature is#    int gsl_sf_bessel_Jn_array (int nmin, int nmax, double x,#                                double result_array[])function sf_bessel_Jn_array(nmin::Integer, nmax::Integer, x::Real)    if nmax < nmin        throw(DomainError())    end    result_array = Vector{Cdouble}(undef, nmax - nmin + 1)    errorcode = @ccall "libgsl".gsl_sf_bessel_Jn_array(                    nmin::Cint, nmax::Cint, x::Cdouble, result_array::Ref{Cdouble})::Cint    if errorcode != 0        error("GSL error code $errorcode")    end    return result_arrayend

The C function wrapped returns an integer error code; the results of the actual evaluation of the Bessel J function populate the Julia arrayresult_array. This variable is declared as aRef{Cdouble}, since its memory is allocated and managed by Julia. The implicit call toBase.cconvert(Ref{Cdouble}, result_array) unpacks the Julia pointer to a Julia array data structure into a form understandable by C.

Fortran Wrapper Example

The following example utilizesccall to call a function in a common Fortran library (libBLAS) to compute a dot product. Notice that the argument mapping is a bit different here than above, as we need to map from Julia to Fortran. On every argument type, we specifyRef orPtr. This mangling convention may be specific to your Fortran compiler and operating system and is likely undocumented. However, wrapping each in aRef (orPtr, where equivalent) is a frequent requirement of Fortran compiler implementations:

function compute_dot(DX::Vector{Float64}, DY::Vector{Float64})    @assert length(DX) == length(DY)    n = length(DX)    incx = incy = 1    product = @ccall "libLAPACK".ddot(        n::Ref{Int32}, DX::Ptr{Float64}, incx::Ref{Int32}, DY::Ptr{Float64}, incy::Ref{Int32})::Float64    return productend

Garbage Collection Safety

When passing data to a@ccall, it is best to avoid using thepointer function. Instead define aBase.cconvert method and pass the variables directly to the@ccall.@ccall automatically arranges that all of its arguments will be preserved from garbage collection until the call returns. If a C API will store a reference to memory allocated by Julia, after the@ccall returns, you must ensure that the object remains visible to the garbage collector. The suggested way to do this is to make a global variable of typeArray{Ref,1} to hold these values until the C library notifies you that it is finished with them.

Whenever you have created a pointer to Julia data, you must ensure the original data exists until you have finished using the pointer. Many methods in Julia such asunsafe_load andString make copies of data instead of taking ownership of the buffer, so that it is safe to free (or alter) the original data without affecting Julia. A notable exception isunsafe_wrap which, for performance reasons, shares (or can be told to take ownership of) the underlying buffer.

The garbage collector does not guarantee any order of finalization. That is, ifa contained a reference tob and botha andb are due for garbage collection, there is no guarantee thatb would be finalized aftera. If proper finalization ofa depends onb being valid, it must be handled in other ways.

Non-constant Function Specifications

In some cases, the exact name or path of the needed library is not known in advance and must be computed at run time. To handle such cases, the library component specification can be a function call, e.g.find_blas().dgemm. The call expression will be executed when theccall itself is executed. However, it is assumed that the library location does not change once it is determined, so the result of the call can be cached and reused. Therefore, the number of times the expression executes is unspecified, and returning different values for multiple calls results in unspecified behavior.

If even more flexibility is needed, it is possible to use computed values as function names by staging througheval as follows:

@eval @ccall "lib".$(string("a", "b"))()::Cint

This expression constructs a name usingstring, then substitutes this name into a new@ccall expression, which is then evaluated. Keep in mind thateval only operates at the top level, so within this expression local variables will not be available (unless their values are substituted with$). For this reason,eval is typically only used to form top-level definitions, for example when wrapping libraries that contain many similar functions. A similar example can be constructed for@cfunction.

However, doing this will also be very slow and leak memory, so you should usually avoid this and instead keep reading. The next section discusses how to use indirect calls to efficiently achieve a similar effect.

Indirect Calls

The first argument to@ccall can also be an expression evaluated at run time. In this case, the expression must evaluate to aPtr, which will be used as the address of the native function to call. This behavior occurs when the first@ccall argument contains references to non-constants, such as local variables, function arguments, or non-constant globals.

For example, you might look up the function viadlsym, then cache it in a shared reference for that session. For example:

macro dlsym(lib, func)    z = Ref{Ptr{Cvoid}}(C_NULL)    quote        let zlocal = $z[]            if zlocal == C_NULL                zlocal = dlsym($(esc(lib))::Ptr{Cvoid}, $(esc(func)))::Ptr{Cvoid}                $z[] = zlocal            end            zlocal        end    endendmylibvar = Libdl.dlopen("mylib")@ccall $(@dlsym(mylibvar, "myfunc"))()::Cvoid

Closure cfunctions

The first argument to@cfunction can be marked with a$, in which case the return value will instead be astruct CFunction which closes over the argument. You must ensure that this return object is kept alive until all uses of it are done. The contents and code at the cfunction pointer will be erased via afinalizer when this reference is dropped and atexit. This is not usually needed, since this functionality is not present in C, but can be useful for dealing with ill-designed APIs which don't provide a separate closure environment parameter.

function qsort(a::Vector{T}, cmp) where T    isbits(T) || throw(ArgumentError("this method can only qsort isbits arrays"))    callback = @cfunction $cmp Cint (Ref{T}, Ref{T})    # Here, `callback` isa Base.CFunction, which will be converted to Ptr{Cvoid}    # (and protected against finalization) by the ccall    @ccall qsort(a::Ptr{T}, length(a)::Csize_t, Base.elsize(a)::Csize_t, callback::Ptr{Cvoid})    # We could instead use:    #    GC.@preserve callback begin    #        use(Base.unsafe_convert(Ptr{Cvoid}, callback))    #    end    # if we needed to use it outside of a `ccall`    return aend
Note

Closure@cfunction relies on LLVM trampolines, which are not available on all platforms (for example ARM and PowerPC).

Closing a Library

It is sometimes useful to close (unload) a library so that it can be reloaded. For instance, when developing C code for use with Julia, one may need to compile, call the C code from Julia, then close the library, make an edit, recompile, and load in the new changes. One can either restart Julia or use theLibdl functions to manage the library explicitly, such as:

lib = Libdl.dlopen("./my_lib.so") # Open the library explicitly.sym = Libdl.dlsym(lib, :my_fcn)   # Get a symbol for the function to call.@ccall $sym(...) # Use the pointer `sym` instead of the library.symbol tuple.Libdl.dlclose(lib) # Close the library explicitly.

Note that when using@ccall with the input (e.g.,@ccall "./my_lib.so".my_fcn(...)::Cvoid), the library is opened implicitly and it may not be explicitly closed.

Variadic function calls

To call variadic C functions asemicolon can be used in the argument list to separate required arguments from variadic arguments. An example with theprintf function is given below:

julia> @ccall printf("%s = %d\n"::Cstring ; "foo"::Cstring, foo::Cint)::Cintfoo = 38

ccall interface

There is another alternative interface to@ccall. This interface is slightly less convenient but it does allow one to specify acalling convention.

The arguments toccall are:

  1. A(:function, "library") pair (most common),

    OR

    a:function name symbol or"function" name string (for symbols in the current process or libc),

    OR

    a function pointer (for example, fromdlsym).

  2. The function's return type

  3. A tuple of input types, corresponding to the function signature. One common mistake is forgetting that a 1-tuple of argument types must be written with a trailing comma.

  4. The actual argument values to be passed to the function, if any; each is a separate parameter.

Note

The(:function, "library") pair, return type, and input types must be literal constants (i.e., they can't be variables, but seeNon-constant Function Specifications).

The remaining parameters are evaluated at compile-time, when the containing method is defined.

A table of translations between the macro and function interfaces is given below.

@ccallccall
@ccall clock()::Int32ccall(:clock, Int32, ())
@ccall f(a::Cint)::Cintccall(:a, Cint, (Cint,), a)
@ccall "mylib".f(a::Cint, b::Cdouble)::Cvoidccall((:f, "mylib"), Cvoid, (Cint, Cdouble), (a, b))
@ccall $fptr.f()::Cvoidccall(fptr, f, Cvoid, ())
@ccall printf("%s = %d\n"::Cstring ; "foo"::Cstring, foo::Cint)::Cint<unavailable>
@ccall printf("%s = %s\n"::Cstring ; "2 + 2"::Cstring, "5"::Cstring)::Cintccall(:printf, Cint, (Cstring, Cstring...), "%s = %s\n", "2 + 2", "5")
<unavailable>ccall(:gethostname, stdcall, Int32, (Ptr{UInt8}, UInt32), hn, length(hn))

Calling Convention

The second argument toccall (immediately preceding return type) can optionally be a calling convention specifier (the@ccall macro currently does not support giving a calling convention). Without any specifier, the platform-default C calling convention is used. Other supported conventions are:stdcall,cdecl,fastcall, andthiscall (no-op on 64-bit Windows). For example (frombase/libc.jl) we see the samegethostnameccall as above, but with the correct signature for Windows:

hn = Vector{UInt8}(undef, 256)err = ccall(:gethostname, stdcall, Int32, (Ptr{UInt8}, UInt32), hn, length(hn))

For more information, please see theLLVM Language Reference.

There is one additional special calling conventionllvmcall, which allows inserting calls to LLVM intrinsics directly. This can be especially useful when targeting unusual platforms such as GPGPUs. For example, forCUDA, we need to be able to read the thread index:

ccall("llvm.nvvm.read.ptx.sreg.tid.x", llvmcall, Int32, ())

As with anyccall, it is essential to get the argument signature exactly correct. Also, note that there is no compatibility layer that ensures the intrinsic makes sense and works on the current target, unlike the equivalent Julia functions exposed byCore.Intrinsics.

Accessing Global Variables

Global variables exported by native libraries can be accessed by name using thecglobal function. The arguments tocglobal are a symbol specification identical to that used byccall, and a type describing the value stored in the variable:

julia> cglobal((:errno, :libc), Int32)Ptr{Int32} @0x00007f418d0816b8

The result is a pointer giving the address of the value. The value can be manipulated through this pointer usingunsafe_load andunsafe_store!.

Note

Thiserrno symbol may not be found in a library named "libc", as this is an implementation detail of your system compiler. Typically standard library symbols should be accessed just by name, allowing the compiler to fill in the correct one. Also, however, theerrno symbol shown in this example is special in most compilers, and so the value seen here is probably not what you expect or want. Compiling the equivalent code in C on any multi-threaded-capable system would typically actually call a different function (via macro preprocessor overloading), and may give a different result than the legacy value printed here.

Accessing Data through a Pointer

The following methods are described as "unsafe" because a bad pointer or type declaration can cause Julia to terminate abruptly.

Given aPtr{T}, the contents of typeT can generally be copied from the referenced memory into a Julia object usingunsafe_load(ptr, [index]). The index argument is optional (default is 1), and follows the Julia-convention of 1-based indexing. This function is intentionally similar to the behavior ofgetindex andsetindex! (e.g.[] access syntax).

The return value will be a new object initialized to contain a copy of the contents of the referenced memory. The referenced memory can safely be freed or released.

IfT isAny, then the memory is assumed to contain a reference to a Julia object (ajl_value_t*), the result will be a reference to this object, and the object will not be copied. You must be careful in this case to ensure that the object was always visible to the garbage collector (pointers do not count, but the new reference does) to ensure the memory is not prematurely freed. Note that if the object was not originally allocated by Julia, the new object will never be finalized by Julia's garbage collector. If thePtr itself is actually ajl_value_t*, it can be converted back to a Julia object reference byunsafe_pointer_to_objref(ptr). (Julia valuesv can be converted tojl_value_t* pointers, asPtr{Cvoid}, by callingpointer_from_objref(v).)

The reverse operation (writing data to aPtr{T}), can be performed usingunsafe_store!(ptr, value, [index]). Currently, this is only supported for primitive types or other pointer-free (isbits) immutable struct types.

Any operation that throws an error is probably currently unimplemented and should be posted as a bug so that it can be resolved.

If the pointer of interest is a plain-data array (primitive type or immutable struct), the functionunsafe_wrap(Array, ptr,dims, own = false) may be more useful. The final parameter should be true if Julia should "take ownership" of the underlying buffer and callfree(ptr) when the returnedArray object is finalized. If theown parameter is omitted or false, the caller must ensure the buffer remains in existence until all access is complete.

Arithmetic on thePtr type in Julia (e.g. using+) does not behave the same as C's pointer arithmetic. Adding an integer to aPtr in Julia always moves the pointer by some number ofbytes, not elements. This way, the address values obtained from pointer arithmetic do not depend on the element types of pointers.

Thread-safety

Some C libraries execute their callbacks from a different thread, and since Julia isn't thread-safe you'll need to take some extra precautions. In particular, you'll need to set up a two-layered system: the C callback should onlyschedule (via Julia's event loop) the execution of your "real" callback. To do this, create anAsyncCondition object andwait on it:

cond = Base.AsyncCondition()wait(cond)

The callback you pass to C should only execute accall to:uv_async_send, passingcond.handle as the argument, taking care to avoid any allocations or other interactions with the Julia runtime.

Note that events may be coalesced, so multiple calls touv_async_send may result in a single wakeup notification to the condition.

More About Callbacks

For more details on how to pass callbacks to C libraries, see thisblog post.

C++

For tools to create C++ bindings, see theCxxWrap package.

  • 1Non-library function calls in both C and Julia can be inlined and thus may have even less overhead than calls to shared library functions. The point above is that the cost of actually doing foreign function call is about the same as doing a call in either native language.
  • 2TheClang package can be used to auto-generate Julia code from a C header file.

Settings


This document was generated withDocumenter.jl version 1.8.0 onWednesday 9 July 2025. Using Julia version 1.11.6.


[8]ページ先頭

©2009-2025 Movatter.jp