C Interoperability¶
This README describes support in the Chapel compiler for referring to C code within Chapel using a keyword named ‘extern’. These features are still in the process of being improved.
External C functions, variables, and types can be referred to within a Chapel program. The section Working with C below describes the basic ideas of how Chapel and C interoperate. There are two supported strategies for providing the Chapel compiler with information about the C declarations that should be usable from Chapel code.
- The explicit strategy is to use extern declarations, as described in the section Support for Extern Declarations below. When using the explicit strategy, one creates a Chapel declarations for each C function, variable, or type that one wants to use from Chapel code. This explicit strategy allows for a great deal of manual control over how the Chapel compiler views a C function or type. This stategy is typically used within the Chapel standard modules for portability reasons.
- The extern block feature provides an implicit strategy, as described in the section Support for Extern Blocks below. This strategy makes use of the clang parser and so requires a Chapel compiler built with LLVM support. The main advantage of the extern block stategy is that it is not necessary to provide Chapel declarations for every desired C function or type. Instead, such declarations are added automatically. In addition, the extern block can contain C function definitions in addition to declarations and so provides a way in which programs that need to be written in both C and Chapel can be implemented in a single source file.
Working with C¶
Whether using extern declarations or extern blocks, it is important to understand the basic interoperability support provided by the Chapel compiler.
The general approach is as follows:
- Programmers inform the Chapel compiler about C concepts that they want to refer to in their Chapel code with explicit extern declarations or by creating extern declarations with extern blocks.
- Programmers use the extern C concepts that have been described. That might include declaring variables, calling functions, or operating on already declared global variables. See Using Extern Declarations.
- Library and header dependencies are described on the Chapel compiler’s command line or with require statements as described in the section Expressing Dependencies.
The next section describes the basic information about how the Chapel compiler views common C types.
Standard C Types¶
In declaring extern C procedures and variables, one of the requirements is to accurately describe their types within the Chapel ‘extern’ declarations. Because the C specification allows compilers to determine how many bits are used in the representation of various types, one must be careful to avoid false assumptions, such as that a C ‘int’ can always be represented via a Chapel ‘int’ or ‘int(32)’.
To help with mapping types between C and Chapel, Chapel installations contain a standard module named SysCTypes (located under $CHPL_HOME/modules/standard/gen/...). This module defines a number of type aliases that accurately describe C types using their Chapel equivalents. Most of these are prefixed by c_ to distinguish them from Chapel type names that may have different meanings.
The Chapel names for C types are:
c_int
c_uint
c_long
c_ulong
c_longlong
c_ulonglong
c_char
c_schar
c_uchar
c_short
c_ushort
ssize_t
size_t
c_void_ptr
c_ptr(T)
c_string
For consistency, the following type aliases are also provided even though their sizes can’t vary in C (thereby permitting the equivalent Chapel types to always be usable):
c_float // (a real(32) in Chapel)
c_double // (a real(64) in Chapel)
c_void_ptr, c_string, and c_ptr(T) are described in the next section.
Pointer and String Types¶
Chapel supports 3 C pointer types: c_void_ptr, c_string, and c_ptr(T). These types are the same as C types:
c_void_ptr is void*
c_ptr(T) is T*
c_string is const char*
Note that in some cases, a ref argument intent may be used in place of c_void_ptr or c_ptr(T).
All three of these pointer types may only point to local memory. The intent is that they will be used to interoperate with C libraries that run within a single locale. In addition, these pointer types must be treated carefully as it is possible to create the same kinds of problems as in C - in particular, it is necessary know that one of these types points to valid memory before it is used. As in C, these pointers could point to a stack variable that is no longer valid, or to memory that has been freed. The Chapel language makes no effort to extend a variable’s lifetime if it is converted in some manner to a C pointer.
c_void_ptr¶
The c_void_ptr type is provided as an opaque C pointer. Since the type is unknown, there is no way to dereference this pointer. In addition, it is not possible to construct a c_void_ptr directly in Chapel. Normally, a C function will return the void pointer, which will be passed to other C functions.
c_ptr(T)¶
The c_ptr(T) type is a generic type representing a C pointer to an arbitrary type T. This pointer should normally only point to local memory - since no communication will be generated when it is dereferenced. Of course, the pointed-to type T should be one that is supported in C interoperability if the c_ptr(T) is used for C interoperability.
ref intents¶
Note that when declaring extern procedures that use function arguments that are passed by pointer in C, it is recommended to use the ref argument intent instead of c_ptr(T). Using the ref intent allows the arguments to the extern proc to be passed directly instead of needing to be converted to a C pointer first. For example, both the functions byRef and byPtr below have the same C prototype, but they must be used differently in Chapel:
// both of these correspond to void fn(int* x)
extern proc byRef(ref x:c_int);
extern proc byPtr(x:c_ptr(c_int));
var x:c_int = 7;
byRef(x); // ref argument intent allows the variable to be passed directly
byPtr(c_ptrTo(x)); // c_ptr argument must be constructed explicitly
c_string¶
The c_string type maps to a constant C string (that is, const char*) that is intended for use locally. A c_string can be obtained from a Chapel string using the method string.c_str(). A Chapel string can be constructed from a C string using the cast operator. Note however that because c_string is a local-only type, the .c_str() method can only be called on Chapel strings that are stored on the same locale; calling .c_str() on a non-local string will result in a runtime error.
Support for Extern Declarations¶
Declaring External C Types¶
You can refer to other external C types using ‘extern’ plus the normal type alias keyword. By leaving off any sort of type definition, you are telling the Chapel compiler that it can’t assume it knows anything about how the type is represented or how to operate on it (though the compiler may make copies of values of the type using assignment and/or memcpy()).
Effectively, such a declaration defines a new primitive type in Chapel. For example, the following declaration says that there is an external type named ‘foo’:
extern type foo;
This permits you to declare variables of type ‘foo’ and to declare external functions that accept or return arguments of type ‘foo’. Because Chapel knows nothing about the type ‘foo’ such variables cannot be manipulated within Chapel apart from passing them between routines. The type symbol ‘foo’ must be declared in a C header file provided on Chapel’s compiler command-line.
If an external C type can be described in Chapel, that definition can be given in Chapel, which permits the compiler to deal with the type directly, as it would any other Chapel type alias. For example, if the external type ‘foo’ was a 64-bit integer, you could describe it in Chapel using:
extern type foo = int(64);
Static, fixed-size C array types can be described within Chapel using its homogeneous tuple type. For example, the following C typedef:
typedef double vec[3];
could be described in Chapel using:
extern type vec = 3*real(64);
To refer to more complex C types like external structs or pointers to structs, see the section on Declaring External C Structs below.
Declaring External C Variables and Constants¶
A C variable or constant can be referred to within Chapel by prefixing its declaration with the extern keyword. For example:
extern var bar: foo;
would tell the Chapel compiler about an external C variable named ‘bar’ of type ‘foo’. Similarly:
extern const baz: int(32);
would refer to an external 32-bit integer constant named ‘baz’ in the C code. In practice, external consts can be used to provide Chapel definitions for #defines and enum symbols in addition to traditional C constants.
Note that it is not currently possible to explicitly declare an extern param. In the future, it might be possible to use an extern block to import #define constants that are known at compile time as param constants within Chapel.
Declaring External C Functions¶
To make a call to an external C function, you will need to prototype the routine in your Chapel code using the extern keyword. For example, for a C function foo() that takes no arguments and returns nothing, the prototype would appear as follows:
extern proc foo();
C functions that return values which you wish to refer to within your Chapel program must have those return types declared. Note that the Chapel compiler will not infer the return type as it does for Chapel functions. To make the function above return a C “double”, it would be declared:
extern proc foo(): real;
Similarly, external functions that expect arguments must declare those arguments in Chapel.
Types of function arguments may be omitted, in which case the types will be inferred based on the Chapel callsite. For example, the following Chapel code:
extern proc foo(x: int, y): real;
var a, b: int;
foo(a, b);
Would imply that external function foo() is able to take two 32-bit integer values and that it returns a 64-bit real (‘double’ in C). External function declarations with omitted type arguments can be used to support calls to external C macros or varargs functions that accept multiple argument signatures.
Default arguments can be declared for external function arguments, in which case the Chapel compiler will supply the default argument value if it is omitted at the callsite. For example:
extern proc foo(x: int, y = 1.2): real;
foo(0);
Would cause external function foo() to be called with the arguments 0 and 1.2.
C varargs functions can be declared using Chapel’s varargs (”...”) syntax. For example, the following declaration prototypes C’s printf function:
extern proc printf(fmt: c_string, vals...?numvals): int;
Note that it can also be prototyped more trivially/less accurately as follows:
extern proc printf(args...): int;
which relies on the callsite to pass in reasonable arguments (otherwise, the C compilation step will likely fail).
External C functions or macros that accept type arguments can also be prototyped in Chapel by declaring the argument as a type. For example:
extern foo(type t);
Calling such a routine with a Chapel type will cause the type identifier (e.g., ‘int’) to be passed to the routine. In practice, this will typically only be useful if the external function is a macro or built-in (like sizeof()) that can handle type identifiers.
Extern functions with array arguments are handled as a special case within the compiler. As an example:
extern proc foo(x: [] int, n: int);
This procedure definition will match up to a C function with the prototype of
void foo(int64_t* x, int64_t n);
The Chapel compiler will then rewrite any calls to foo like this:
foo(x, 10); // -> foo(c_ptrTo(x), 10);
Note that this same technique won’t work for distributed rectangular arrays, nor for associative, sparse, or opaque arrays because their data isn’t necessarily stored using a representation that translates trivially to a C array.
It is possible to provide the Chapel compiler with a different name for the function than the name available to other Chapel code. For example, if there is a C function called foo_in_c returning an int, we might use the following declaration to provide that function to other Chapel code with the name foo_in_chapel
extern "foo_in_c" proc foo_in_chapel(): c_int;
writeln(foo_in_chapel()); // will generate a call to foo_in_c
Frequently Asked Questions About Declaring External Routines¶
How do I pass a Chapel variable to an external routine that expects a pointer?
Today, the preferred way to do this is to declare the argument as having ‘ref’ intent. This should cause the Chapel compiler to pass a pointer to the argument. For example, given a C function:
void foo(double* x);
This could be called in Chapel using:
extern proc foo(ref x: real);
var myVal: real = 1.2;
foo(myVal);
Declaring External C Structs¶
External C struct types can be referred to within Chapel by prefixing a Chapel record definition with the extern keyword. For example, given an external C structure defined in foo.h called fltdbl:
typedef struct _fltdbl {
float x;
double y;
} fltdbl;
This type could be referred to within a Chapel program using:
extern record fltdbl {
}
Within the Chapel declaration, some or all of the fields from the C structure may be specified. Any fields that are not specified (or that cannot be specified because there is no equivalent Chapel type) cannot be referenced within the Chapel code. The order of these fields need not match the order they were specified within the C code. As an example, the following declaration would permit access to both fields x and y within variables of type fltdbl:
extern record fltdbl {
var x: real(32);
var y: real(64);
}
as would the following declaration:
extern record fltdbl {
var y: real(64);
var x: real(32);
}
Alternatively, the external declaration could only mention one of the fields. For example, the following declaration would permit field y to be accessed by a Chapel program, but not field x (though both would still be stored in any variable of type ‘fltdbl’).
extern record fltdbl {
var y: real(64);
}
Alternatively, the external declaration can avoid mentioning any fields, which would permit a variable of that struct type to be passed between Chapel and C routines, but without permitting its fields to be accessed within the Chapel program:
extern record fltdbl {
}
A C header file containing the struct’s definition in C must be specified on the chpl compiler command line or in a require statement as described in Expressing Dependencies.
To work with a C structure that is not typedef’d, just name the C type name after the extern keyword. In the example below, the struct stat type does not have a corresponding typedef in C. Therefore, we inform the Chapel compiler that the C name for the type is struct stat:
extern "struct stat" record chpl_stat_type {
var st_size: off_t;
}
proc getFileSize(path:c_string) : int {
extern proc stat(x: c_string, ref buf:chpl_stat_type): c_int;
var buf: chpl_stat_type;
if (chpl_stat_function(path, buf) == 0) {
return buf.st_size;
}
return -1;
}
writeln(getFileSize("stat-example.chpl"));
Note that external record types only support assignment from records of matching type. In particular, Chapel’s normal mechanisms that perform record assignment by field name are not used for external records. This restriction could be lifted in the future if considered useful to users.
Referring to External C Pointer-to-Structs (“extern classes”)¶
You can also refer to an external C pointer-to-struct types by considering it to be an ‘extern class’ in Chapel. The declaration style is similar to that described above, it simply has different implications on the underlying C types.
As an example, given the declaration:
extern class D {
var x: real;
}
The requirements on the corresponding C code are:
- There must be a struct type that is typedef’d to have the name _D.
- A pointer-to-_D type must be typedef’d to have the name D.
- The _D struct type must contain a field named ‘x’ of type double. Like external records/structs, it may also contain other fields that will simply be ignored by the Chapel compiler.
Thus, the following C typedef would fulfill the external Chapel class declaration shown above:
typedef struct __D {
double x;
int y;
} _D, *D;
where the Chapel compiler would not know about the ‘y’ field and therefore could not refer to it or manipulate it.
Opaque Types¶
NOTE: This support may eventually be deprecated as other ‘extern’ features become increasingly flexible and robust.
You can refer to other external pointer-based C types that cannot be described in Chapel using the “opaque” keyword. As the name implies, these types are opaque as far as Chapel is concerned and cannot be used for operations other than argument passing and assignment (to/from other similarly opaque types).
For example, Chapel could be used to call an external C function that returns a pointer to a structure (that we can’t or won’t describe as an external class using the previous mechanism) as follows:
extern proc returnStructPtr(): opaque;
var structPtr: opaque = returnStructPtr();
However, because structPtr’s type is opaque, it cannot be used for much apart from assigning it to other opaque variables of matching underlying type, or passing it back to an external C routine that expects a pointer-to-struct of that type:
extern proc operateOnStructPtr(ptr: opaque);
var copyOfStructPtr = structPtr;
operateOnStructPtr(structPtr);
Support for Extern Blocks¶
[Note: The features in this section rely on Chapel to being built with llvm/clang enabled. To do so, set environment variable CHPL_LLVM to ‘llvm’ and rebuild your Chapel installation. See LLVM Support.].
C code and header files can be included directly within Chapel source code using an extern block as follows:
extern {
#include "my_C_API.h"
static int myCInt = 7;
....
}
Such ‘extern { }’ block statements add the top-level C statements to the enclosing Chapel module. This is similar to what one might do manually using the extern declarations (as described above), but can save a lot of labor for a large API. Moreover, using an inline extern block permits you to write C declarations directly within Chapel without having to create distinct C files.
If you don’t want to have a lot of C symbols cluttering up a module’s namespace, it’s easy to put the C code into its own Chapel module:
module C {
extern {
static int foo(int x) { return x + 1; }
... c code here...
}
}
writeln(C.foo(3));
This feature strives to support C global variables, functions, structures, typedefs, enums, and some #defines. Structures always generate a Chapel record, and pointers to a structure are represented with c_ptr(struct type). Also, pointer arguments to functions are always represented with c_ptr or c_string instead of the ref intent.
Note that functions or variables declared within an extern block should either be declared in a .h file that is #included or they should be be declared static (otherwise you might get a warning during C compilation).
Also note that a static inline function can be inlined into Chapel code that calls it. Thus, by using static inline in an extern block, it is possible to hand-tune a computational kernel by writing some of it in inline assembly.
#defines¶
The extern block functionality allows simple #defines to be used from Chapel code. Simple defines take no arguments and define an integer value or use another #define. For example:
extern {
#define NEG_ONE (-1)
#define MY_NUMBER (NEG_ONE)
}
writeln(MY_NUMBER);
However, it is easy to create #defines that are not supported because the Chapel compiler is unable to determine their expression type. In particular, any #define taking arguments is not currently supported. For example, the ADD definition below has a resulting expression type that cannot be determined by the Chapel compiler:
extern {
#define ADD(x,y) ((x)+(y)) // NOT SUPPORTED
}
var x = ADD(1,2); // ERROR - ADD not defined in Chapel.
If the library you are using depends on these types, the current solution is to create inline functions to call the define with the appropriate argument types; for example:
extern {
#define ADD(x,y) ((x)+(y))
static inline int ADD_int_int(int x, int y) { return ADD(x,y); }
}
var sum = ADD_int_int(1,2);
writeln(sum);
Pointer Types¶
See the section Pointer and String Types above for background on how the Chapel programs can work with C pointer types. Any pointer type used in an extern block will be made visible to the Chapel program as c_ptr(T) or c_string (for const char* types).
For example:
extern {
static void setItToOne(int* x) { *x = 1; }
// will translate automatically into
// extern proc setItToOne(x:c_ptr(c_int));
// The Chapel compiler can't know if X is used as an array,
// if the argument will come from a Chapel variable, and in more general
// cases, the best way to handle multiple levels of pointers. For example:
static void setSpace(int** x) {
static int space[10];
*x = space;
}
// translates automatically into
// extern proc returnSpace( x:c_ptr(c_ptr(c_int)) );
static void setString(const char** x) { *x = "My String"; }
}
var x:c_int;
setItToOne(c_ptrTo(x));
var space:c_ptr(c_int);
setSpace(c_ptrTo(space));
var str:c_string;
setString(c_ptrTo(str));
writeln(toString(str));
As you can see in this example, using the extern block might result in more calls to c_ptrTo() when using the generated extern declarations, because the compiler cannot automatically distinguish between several common cases (passing an array vs. passing in an argument by reference).
Example¶
Here’s a more complete example:
module MyCModule {
extern {
static int foo(int b) { return b + 1; }
}
}
writeln("Hello");
writeln(MyCModule.foo(7));
This prints out
Hello
8
which demonstrates full integration between Chapel and the C function it calls.
Using Extern Declarations¶
Extern declarations can be used just like normal Chapel types, variables, or functions. Using these extern declarations from Chapel code requires some care.
Parallel Safety¶
If external routines are to be called from Chapel within parallel execution contexts, they should be parallel-safe. As with internal Chapel routines, it is also the responsibility of the Chapel programmer to call external routines in a manner that preserves the integrity of objects accessible to those routines. Simply put, objects shared across Chapel tasks must be kept parallel-safe within Chapel.
Multiple Locales¶
Since the extern C code does not generally have any support for multiple locales, it is important to take care when using this code from multiple locales. Here are some things to be aware of:
- extern global variables will refer to a local version of that variable on each locale. These variables might become different if they are changed differently on different locales.
- c_ptr is always generated as a narrow pointer type (in other words, it does not encode the locale storing the pointed-to value - just an address). That means that passing a c_ptr from one locale to another and then using it on the second locale will probably result in a core dump.
Initialization¶
Chapel variables of extern type are not generally initialized automatically. Be sure to manually initialize Chapel variables of extern type.
In the future, we would like to support automatic zero initialization of such variables and a way to provide their default initializer.
Working with c_ptr¶
A c_ptr type can be constructed from a Chapel variable using the c_ptrTo() function; for example:
var i:c_int;
var i_ptr = c_ptrTo(i); // now i_ptr has type c_ptr(c_int) == int* in C
Since a C pointer might refer to a single variable or an array, the c_ptr type supports 0-based array indexing and dereferencing. In addition, it is possible to allocate and free space for one or more elements and return the result as a c_ptr. See the following example:
var cArray = c_calloc(c_int, 10);
for i in 0..#10 {
cArray[i] = i:c_int;
}
// c_ptr.deref() always refers to the first element.
cArray.deref() = 17;
for i in 0..#10 {
writeln(cArray[i]);
}
c_free(cArray);
Variables of type c_ptr can be compared against or set to nil.
Working with strings¶
If you need to call a C function and provide a Chapel string, you may need to convert the Chapel string to a C string first. Chapel string literals will automatically convert to C strings. A Chapel string variable can be converted using the string.c_str method.
myprint.h:
void myprint(const char* str);
myprint.c:
void myprint(const char* str) {
printf("%s\n", str);
}
myprint.chpl:
extern proc myprint(str:c_string);
// string literal is automatically converted to a c_string
myprint("hello");
// a string variable must be converted with .c_str()
var s = "goodbye";
myprint(s.c_str());
Expressing Dependencies¶
Any required C header files, source code files, object files, or library files must be provided to the Chapel compiler by one of two mechanisms.
They can be listed at compile-time on the Chapel command line For example, if an external function foo() was defined in foo.h and foo.c, it could be added to the compilation using any of these commands:
chpl foo.h foo.c myProgram.chpl chpl foo.h foo.o myProgram.chpl #if foo.c had already been compiled) chpl foo.h -lfoo myProgram.chpl #if foo.c had been archived in libfoo.a)Note that you can use -I and -L arguments for the Chapel compiler to specify include or library paths as with a C compiler.
Alternatively, the required C resources can be listed within the Chapel file using the require statement. For example:
require "foo.h", "foo.c";This has an effect similar to adding foo.h and foo.c to the command line. You might need to specify -I and -L arguments to indicate to the directories storing any headers or library files.
Similarly, the version below uses the require statement to indicate that this module depends on libfoo.a (and has a similar effect as if -lfoo were added to the command line).
require "foo.h", "-lfoo";
Either approach has the following results:
- During Chapel’s C code generation stage, any header files listed on the compiler’s command line or in a require statement will be #include’d by the generated code in order to ensure that the appropriate prototypes are found before making any references to the external symbols.
- During Chapel’s C compilation stage, any C files on the command line or in a require statement will be compiled using the same flags as the Chapel-generated C files (use –print-commands to see these compile commands).
- During Chapel’s link step, any .o and .a files listed on the compiler’s command-line or in require statements will be included in the final executable.
Future Directions¶
We intend to continue improving these capabilities to provide richer and richer support for external types and functions over time. If you have specific requests for improvement, please let us know at: chapel_info@cray.com.