SWIG Internals Manual

$Header$

(Note : This is a work in progress.)

Table of Contents

• 1. Introduction
  • 1.1 Directory Guide
  • 1.2 Overall Program Flow
• 2. DOH
  • 2.1 Motivation and Background
  • 2.2 Basic Types
  • 2.3 Creating, Copying and Destroying Objects
  • 2.4 A Word About Mutability and Copying
  • 2.5 Strings
  • 2.6 Lists
  • 2.7 Hash Tables
  • 2.8 Files
  • 2.9 Void Objects
  • 2.10 Utility Functions
• 3. Types and Typemaps
• 4. Parsing
• 5. Difference Between SWIG 1.1 and SWIG 1.3
• 6. Plans for SWIG 2.0
• 7. C/C++ Wrapper Support Functions
• 8. Reserved
• 9. Reserved
• 10. Guile Support
  • 10.1 Meaning of "Module"
  • 10.2 Linkage
  • 10.3 Underscore Folding
  • 10.4 Typemaps
  • 10.5 Smobs
  • 10.5 Exception Handling
• 11. Python Support
• 12. Perl Support
• 13. Java Support

The audience is assumed to be SWIG developers (who should also read the SWIG Engineering Manual before starting to code).

1.1 Directory Guide

• Modules1.1/swigmain.cxx:main() is the program entry point. It parses the language-specifying command-line option (for example, -java), creating a new language-specific wrapping object (each language is a C++ class derived from base class Language). This object and the command-line is passed to SWIG_main(), whose return value is the program exit value.
• SWIG1.1/main.cxx:SWIG_main() is the "real" main. It initializes the preprocessor and typemap machinery, defines some preprocessor symbols, locates the SWIG library, processes common command-line options, and then calls the language-specific command-line parser. From here there are three paths: "help", "checkout" and everything else.
  • In "help" mode, clean up open files and exit.
  • In "checkout" mode, copy specified files from the SWIG library to the current directory. Errors cause error messages but no non-lcoal exits.
  • Otherwise, do wrapping: determine output file name(s), define some preprocessor symbols and run the preprocessor, initialize the interface-definition parser, set up the typemap for handling new return strings, and finally do the language-specific parse (by calling the language object's parse() method), which creates output files by side-effect.
  Afterwards, remove temporary files, and clean up. If the command-line included -freeze, go into an infinite loop; otherwise return the error count.
• The language-specific parse() (and all other language-specific code) lives in Modules1.1/foo.{h,cxx} for language Foo. Typically, FOO::parse() calls FOO::headers() and then the global function yyparse(), which uses the callbacks registered by SWIG_main() above.

DOH takes a different approach to tackling the complexity problem. First, rather than going overboard with dozens of types and class definitions, DOH only defines a handful of simple yet very useful objects that are easy to remember. Second, DOH uses dynamic typing---one of the features that make scripting languages so useful and which make it possible to accomplish things with much less code. Finally, DOH utilizes a few coding tricks that allow it to perform a limited form of function overloading for certain C datatypes (more on that a little later).

The key point to using DOH is that instead of thinking about code in terms of highly specialized C data structures, just about everything ends up being represented in terms of a just a few datatypes. For example, structures are replaced by DOH hash tables whereas arrays are replaced by DOH lists. At first, this is probably a little strange to most C/C++ programmers, but in the long run in makes the system extremely flexible and highly extensible. Also, in terms of coding, many of the newly DOH-based subsystems are less than half the size (in lines of code) of the earlier C++ implementation.

2.2 Basic Types

• String. A string of characters with automatic memory management and high-level operations such as string replacement. In addition, strings support file I/O operations that make it possible to use them just about anyplace a file can be used.

• List. A list of arbitrary DOH objects (of possibly mixed types).

• Hash. A hash table that maps a set of string keys to a set of arbitrary DOH objects. The DOH version of an associative array for all of you Perl fans.

• File. A DOH wrapper around the C FILE * structure. This is provided since other objects sometimes want to behave like files (strings for instance).

• Void. A DOH wrapper around an arbitrary C pointer. This can be used if you want to place arbitrary C data structures in DOH lists and hash tables.

• DOHString *. A String object.
• DOHList *. A list object.
• DOHHash *. A hash object.
• DOHFile *. A file object.
• DOHVoid *. A void object.
• DOHString_or_char *. A DOH String object or a raw C "char *".

• NewString(DOHString_or_char *value)
  Create a new string object with contents initially set to value. value can be either a C string or a DOH string object.

• NewStringf(char *fmt, ...)
  Create a new string object with contents initially set to a formatted string. Think of this as being sprintf() combined with an object constructor.

• NewList()
  Create a new list object that is initially empty.

• NewHash()
  Create a new hash object that is initially empty.

• NewFile(DOHString_or_char *filename, char *mode)
  Open a file and return a file object. This is a wrapper around the C fopen() library call.

• NewFileFromFile(FILE *f)
  Create a new file object given an already opened FILE * object.

• NewVoid(void *obj, void (*del)(void *))
  Create a new DOH object that is a wrapper around an arbitrary C pointer. del is an optional destructor function that will be called when the object is destroyed.

DOH *a, *b, *c, *d;
a = NewString("Hello World");
b = NewList();
c = Copy(a);         /* Copy the string a */
d = Copy(b);         /* Copy the list b */

Objects can be deleted using the Delete() function. For example:

DOH *a = NewString("Hello World");
...
Delete(a);              /* Destroy a */

DOH *a, *b;
a = NewString("Hello World");
b = NewList();
Append(b,a);         /* Increases refcnt of a to 2 */
Delete(a);           /* Decreases refcnt of a to 1 */
Delete(b);           /* Destroys b, and destroys a */

DOH *a = NewString("Hello");
DohIncref(a);

The representation and manipulation of types is currently in the process of being reorganized and (hopefully) simplified. The following list describes the current set of functions that are used to manipulate datatypes. These functions are different than in SWIG1.1 and may change names in the final SWIG1.3 release.

• DataType_str(DataType *t, char *name).
  This function produces the exact string representation of the datatype t. name is an optional parameter that specifies a declaration name. This is used when dealing with more complicated datatypes such as arrays and pointers to functions where the output might look something like "int (*name)(int, double)".

• DataType_lstr(DataType *t, char *name).
  This function produces a string representation of a datatype that can be safely be assigned a value (i.e., can be used as the "lvalue" of an expression). To do this, qualifiers such as "const", arrays, and references are stripped away or converted into pointers. For example:

  Original Datatype              lstr()
  ------------------             --------
  const char *a                  char *a
  double a[20]                   double *a
  double a[20][30]               double *a
  double &a                      double *a
  

  The intent of the lstr() function is to produce local variables inside wrapper functions--all of which must be reassignable types since they are the targets of conversions from a scripting representation.

• DataType_rcaststr(DataType *t, char *name).
  This function produces a string that casts a type produced by the lstr() function to the type produced by the str() function. You might view it as the inverse of lstr(). This function only produces output when it needs to (when str() and lstr() produce different results). Furthermore, an optional name can be supplied when the cast is to be applied to a specific name. Examples:

  Original Datatype             rcaststr()
  ------------------            ---------
  char *a                       
  const char *a                 (const char *) name
  double a[20]                  (double *) name
  double a[20][30]              (double (*)[30]) name
  double &a                     (double &) *name
  

• DataType_lcaststr(DataType *t, char *name).
  This function produces a string that casts a type produced by the str() function to the type produced by the lstr() function. This function only produces output when it needs to (when str() and lstr() produce different results). Furthermore, an optional name can be supplied when the cast is to be applied to a specific name.

  Original Datatype             lcaststr()
  ------------------            ---------
  char *a                       
  const char *a                 (char *) name
  double a[20]                  (double *) name
  double a[20][30]              (double *) name
  double &a                     (double *) &name
  

• DataType_manglestr(DataType *t).
  Produces a type-string that is used to identify this datatype in the target scripting language. Usually this string looks something like "_double_pp" although the target language may redefine the output for its own purposes. Normally this function strips all qualifiers, references, and arrays---producing a mangled version of the type produced by the lstr() function.

int foo(int a, double b[20][30], const char *c, double &d);

wrapper_foo() {
   lstr("int","result")
   lstr("int","arg0")
   lstr("double [20][30]", "arg1")
   lstr("const char *", "arg2")
   lstr("double &", "arg3")
   ...
   get arguments
   ...
   result = (lcaststr("int"))  foo(rcaststr("int","arg0"),
                               rcaststr("double [20][30]","arg1"),
                               rcaststr("const char *", "arg2"),
                               rcaststr("double &", "arg3"))
   ...
}

wrapper_foo() {
   int      result;
   int      arg0;
   double  *arg1;
   char    *arg2;
   double  *arg3;
   ...
   get arguments
   ...
   result = (int) foo(arg0,
                      (double (*)[30]) arg1,
                      (const char *) arg2,
                      (double &) *arg3);
   ...
}

• For convenience, the string generation functions return a "char *" that points to statically allocated memory living inside the type library. Therefore, it is never necessary (and it's an error) to free the pointer returned by the functions. Also, if you need to save the result, you should make a copy of it. However, with that said, it is probably worth nothing that these functions do cache the last 8 results. Therefore, it's fairly safe to make a handful of repeated calls without making any copies.

When SWIG generates wrappers, it tries to provide a mostly seamless integration with the original code. However, there are a number of problematic features of C/C++ programs that complicate this interface.

• Passing and returning structures by value. When used, SWIG converts all pass-by-value functions into wrappers that pass by reference. For example:

  double dot_product(Vector a, Vector b);
  

  gets turned into a wrapper like this:

  double wrap_dot_product(Vector *a, Vector *b) {
       return dot_product(*a,*b);
  }
  

  Functions that return by value require a memory allocation to store the result. For example:

  Vector cross_product(Vector *a, Vector *b);
  

  become

  Vector *wrap_cross_product(Vector *a, Vector *b) {
     Vector *result = (Vector *) malloc(sizeof(Vector));
     *result = cross_product(a,b);
     return result;
  }
  

  Note: If C++ is being wrapped, the default copy constructor is used instead of malloc() to create a copy of the return result.

• C++ references. C++ references are handled exactly the same as pass/return by value except that a memory allocation is not made for functions that return a reference.

• Qualifiers such as "const" and "volatile". SWIG strips all qualifiers from the interface presented to the target language. Besides, what in the heck is "const" in Perl anyways?

• Instance Methods. Method invocations are handled as a function call in which a pointer to the object (the "this" pointer) appears as the first argument. For example, in the following class:

  class Foo {
  public:
      double bar(double);
  };
  

  The "bar" method is wrapped by a function like this:

  double Foo_bar(Foo *self, double arg0) {
     return self->bar(arg0);
  }
  

• Structure/class data members. Data members are handled by creating a pair of wrapper functions that set and get the value respectively. For example:

  struct Foo {
      int x;
  };
  

  gets wrapped as follows:

  int Foo_x_get(Foo *self) {
      return self->x;
  }
  int Foo_x_set(Foo *self, int value) {
      return (self->x = value);
  }
  

• Constructors. Constructors for C/C++ data structures are wrapped by a function like this:

  Foo *new_Foo() {
      return new Foo;
  }
  

  Note: For C, new objects are created using the calloc() function.

• Destructors. Destructors for C/C++ data structures are wrapper like this:

  void delete_Foo(Foo *self) {
      delete self;
  }
  

  Note: For C, objects are destroyed using free().

• char *Swig_clocal(DataType *t, char *name, char *value)
  This function creates a string containing the declaration of a local variable with type t, name name, and default value value. This local variable is stripped of all qualifiers and will be a pointer if the type is a reference or user defined type.

• DataType *Swig_clocal_type(DataType *t)
  Returns a type object corresponding to the type string produced by the Swig_clocal() function.

• char *Swig_clocal_deref(DataType *t, char *name)
  This function is the inverse of the clocal() function. Given a type and a name, it produces a string containing the code needed to cast/convert the type produced by Swig_clocal() back into it's original type.

• char *Swig_clocal_assign(DataType *t, char *name)
  Given a type and name, this produces a string containing the code (and an optional cast) needed to make an assignment from the real datatype to the local datatype produced by Swig_clocal(). Kind of the opposite of deref().

• int Swig_cargs(Wrapper *w, ParmList *l)
  Given a wrapper function object and a list of parameters, this function declares a set of local variables for holding all of the parameter values (using Swig_clocal()). Returns the number of parameters. In addition, this function sets the local name of each parameter which can be retrieved using the Parm_Getlname() function.

• void Swig_cresult(Wrapper *w, DataType *t, char *resultname, char *decl)
  Generates the code needed to set the result of a wrapper function and performs all of the needed memory allocations for ANSI C (if necessary). t is the type of the result, resultname is the name of the result variable, and decl is a string that contains the C code which produces the result.

• void Swig_cppresult(Wrapper *w, DataType *t, char *resultname, char *decl)
  Generates the code needed to set the result of a wrapper function and performs all of the needed memory allocations for C++ (if necessary). t is the type of the result, resultname is the name of the result variable, and decl is a string that contains the C code which produces the result.

• Wrapper *Swig_cfunction_wrapper(char *fname, DataType *rtype, ParmList *parms, char *code)
  Create a wrapper around a normal function declaration. fname is the name of the wrapper, rtype is the return type, parms are the function parameters, and code is a string containing the code in the function body.

• Wrapper *Swig_cmethod_wrapper(char *classname, char *methodname, DataType *rtype, DataType *parms, char *code)
  

• char *Swig_cfunction_call(char *name, ParmList *parms) This function produces a string containing the code needed to call a C function. The string that is produced contains all of the transformations needed to convert pass-by-value into pass-by-reference as well as handle C++ references. Produces a string like "name(arg0, arg1, ..., argn)".

double dot_product(Vector a, Vector b);

ParmList *l = ... parameter list of the function ...
DataType *t = ... return type of the function ...
char     *name = ... name of the function ...
Wrapper *w = NewWrapper();
Printf(w->def,"void wrap_%s() {\n", name);

/* Declare all of the local variables */
Swig_cargs(w, l);

/* Convert all of the arguments */
...

/* Make the function call and declare the result variable */
Swig_cresult(w,t,"result",Swig_cfunction(name,l));

/* Convert the result into whatever */
...

Printf(w->code,"}\n");
Wrapper_print(w,out);

void wrap_dot_product() {
    Vector *arg0;
    Vector *arg1;
    double  result;

    ...
    result = dot_product(*arg0, *arg1);
    ...
}

Notes:

• The intent of these functions is to provide consistent handling of function parameters and return values so that language module writers don't have to worry about it too much.

• These functions may be superceded by features in the new typemap system which provide hooks for specifying local variable declarations and argument conversions.

This section details guile-specific support in SWIG.

10.1 Meaning of "Module"

There are three different concepts of "module" involved, defined separately for SWIG, Guile, and Libtool. To avoid horrible confusion, we explicitly prefix the context, e.g., "guile-module".

10.2 Linkage

Guile support is complicated by a lack of user community cohesiveness, which manifests in multiple shared-library usage conventions. A set of policies implementing a usage convention is called a linkage. The default linkage is the simplest; nothing special is done. In this case SWIG_init() is provided and users must do something like this:

(define my-so (dynamic-link "./example.so"))
(dynamic-call "SWIG_init" my-so)

A second linkage creates "libtool dl module" wrappers, and currently is broken. Whoever fixes this needs to track Guile's libtool dl module convention, since that is not finalized.

The only other linkage supported at this time creates shared object libraries suitable for use by hobbit's (hobbit4d link) guile module. This is called the "hobbit" linkage, and requires also using the "-package" command line option to set the part of the module name before the last symbol. For example, both command lines: [checkme:ttn]

swig -guile -package my/lib foo.i
swig -guile -package my/lib -module foo foo.i

There are no other linkage types planned, but that could change... To add a new type, add the name to the enum in guile.h and add the case to GUILE::emit_linkage().

10.3 Underscore Folding

Underscores are converted to dashes in identifiers. Guile support may grow an option to inhibit this folding in the future, but no one has complained so far.

10.4 Typemaps

It used to be that the mappings for "native" types were included in guile.cxx. This information is now in Lib/guile/typemaps.i, which presents a new challenge: how to have SWIG include typemaps.i before processing the user's foo.i. At this time, we must say:

%include guile/typemaps.i

For pointer types, SWIG uses Guile smobs. This feature used to be available only SWIG was invoked with the command-line option "-with-smobs", but as of 2000/07/06 (swig-1.3a4) this is the default behavior; the command-line option "-with-smobs" is ignored silently. We plan to drop silent recognition of this option after 1.3a4.

Currently, one wrapper module must be generated without -c and compiled with -DSWIG_GLOBAL, all the other wrapper modules must be generated with -c. Maybe one should move all the global helper functions that come from guile.swg into a library, which is built by make runtime.

In earlier versions of SWIG, C pointers were represented as Scheme strings containing a hexadecimal rendering of the pointer value and a mangled type name. As Guile allows registering user types, so-called "smobs" (small objects), a much cleaner representation has been implemented now. The details will be discussed in the following.

A smob is a cons cell where the lower half of the CAR contains the smob type tag, while the upper half of the CAR and the whole CDR are available. SWIG_Guile_Init() registers a smob type named "swig" with Guile; its type tag is stored in the variable swig_tag. The upper half of the CAR store an index into a table of all C pointer types seen so far, to which new types seen are appended. The CDR stores the pointer value. SWIG smobs print like this: #<swig struct xyzzy * 0x1234affe> Two of them are equal? if and only if they have the same type and value.

To construct a Scheme object from a C pointer, the wrapper code calls the function SWIG_Guile_MakePtr_Str(), passing both a mangled type string and a pretty type string. The former is looked up in the type table to get the type index to store in the upper half of the CAR. If the type is new, it is appended to type table.

To get the pointer represented by a smob, the wrapper code calls the function SWIG_Guile_GetPtr_Str, passing the mangled name of the expected pointer type, which is used for looking up the type in the type table and accessing the list of compatible types. If the Scheme object passed was not a SWIG smob representing a compatible pointer, a wrong-type-arg exception is raised.

10.6 Exception Handling

SWIG code calls scm_error on exception, using the following mapping:

      MAP(SWIG_MemoryError,	"swig-memory-error");
      MAP(SWIG_IOError,		"swig-io-error");
      MAP(SWIG_RuntimeError,	"swig-runtime-error");
      MAP(SWIG_IndexError,	"swig-index-error");
      MAP(SWIG_TypeError,	"swig-type-error");
      MAP(SWIG_DivisionByZero,	"swig-division-by-zero");
      MAP(SWIG_OverflowError,	"swig-overflow-error");
      MAP(SWIG_SyntaxError,	"swig-syntax-error");
      MAP(SWIG_ValueError,	"swig-value-error");
      MAP(SWIG_SystemError,	"swig-system-error");

The default when not specified here is to use "swig-error". See Lib/exception.i for details.

10. Python Support