8. Virtual Machine Reference

8.2. Startup options

The following command-line options are recognised by the Virtual Machine:

Argument Description
-f filename Reads filename (which should contain Self source) immediately after startup (after reading the snapshot) and evaluates the contents. Useful for setting options, installing personal shortcuts, etc.
-h Prints a message describing the options.
-p Suppresses execution of the expression snapshotAction postRead after reading a snapshot. Useful if something in the startup sequence causes the system to break.
-s snapshot Reads initial world from snapshot. A snapshot begins with the line exec Self -s $0 $@ which causes the Virtual Machine to begin execution with the snapshot.
-w Don’t print warnings about object code.

These options are provided for use by Self VM implementors:

Argument Description
-F Discards any machine code saved in the snapshot. If the code in a snapshot is for some reason corrupted, but the objects are not, this option can be used to recover the snapshot.
-l logfile Writes a log of events generated by the spy to logfile.
-r Disables real timer interrupts.
-t Disables all timers.

Other command-line options are ignored by the Virtual Machine but are available at Self level via the primitive _CommandLine.

The standard set of Self objects (built by the worldBuilder.self script) also defines -b (where the objects director is) and -o (for specifying build options)

8.3. System-triggered messages

Certain events cause the system to automatically send a message to the lobby. After reading a snapshot the expression snapshotAction postRead is evaluated. This allows the Self world to reinitialize itself — for example, to reopen windows.

There are other situations in which the system sends messages; see Run-time message lookup errors.

8.4. Run-time message lookup errors

If an error occurs during a message send, the system sends a message to the receiver of the message. Any object can handle these errors by defining (or inheriting) a slot with the corresponding selector. All messages sent by the system in response to a message lookup error have the same arguments. The first argument is the offending message’s selector; the additional arguments specify the message send type (one of ’normal’, ’implicitSelf’, ’undirectedResend’, ’directedResend’, or ’delegated’), the directed resend parent name or the delegatee (0 if not applicable), the sending method holder, and a vector containing the arguments to the message, if any.

The receiver does not understand the message: no slot matching the selector can be found in the receiver or its ancestors.
There is more than one slot matching the selector.
The parent slot through which the resend should have been directed was not found in the sending method holder.
The number of arguments supplied to the _Perform primitive does not match the number of arguments required by the selector.
The first argument to the _Perform primitive (the selector) wasn’t a canonical string.

These error messages are just like any other message. Therefore, it is possible that the object P causing the error (which is being sent the appropriate error message) does not understand the error message M either. If this happens, the system sends the first message (undefinedSelector:) to the current process, with the error message M as argument. If this is not understood, then the system suspends the process. If the scheduler is running, it is notified of the failure.

The system will also suspend a process if it runs out of stack space (too much recursion) or if a block is evaluated whose lexically-enclosing scope has already returned. Since these errors are nonrecoverable they cannot be caught by the same Self process; the scheduler, if running, is notified.

8.5. Low-level error messages

Five kinds of errors can occur during the execution of a Self program: lookup errors, primitive errors, programmer defined errors, non-recoverable errors, and fatal VM errors. All but the last of these are usually caught and handled by mechanisms in the programming environment, resulting in a debugger being presented to the user. However, if programs are run without the programming environment, or the error-handling mechanisms themselves are broken, low-level error facilities are used.

This section describes the various error messages presented by the low-level facilities. For each category or error, the general layout of error messages in that category will be explained along with the format of the stack trace. Then a “rogue’s gallery” of the errors in that category will be shown.

By default, errors are handled by a set of methods defined in module errorHandling. For all errors except nonrecoverable and fatal VM errors, an object can handle errors in its own way by defining its own error handling methods. If the object in which an error occurs neither inherits nor defines error handling behavior, the VM prints out a low-level error message and a stack trace. The system will also resort to this low-level message and trace if an error is encountered while trying to handle an error.

8.6. An example

Here is an expression that produces an error in the current system:

“Self 7” 100000 factorial
The stack has grown too big.
(Self limits stack sizes, and cannot resume processes with stack overflows.)
To debug type “attach” or to show stack type “zombies first printError”.

The error arose because the recursive method factorial exceeded the size allocated for the process stack which resulted in a stack overflow.

The virtual machine currently allocates a fixed-size stack to each process and does not extend the stack on demand.

8.7. Lookup errors

Lookup errors occur when an object does not understand a message that is sent to it. How the actual message lookup is done is described in the Language Reference chapter.

No ’foo’ slot found in shell <0>.
The lookup found no slot matching the selector foo.
No ’fish’ delegatee slot was found in <a child of lobby> <12>.
The lookup found no parent slot fish, which was explicitly specified as the delegatee of the message.
More than one ’system’ slot was found in shell <0>.
The matching slots are: oddballs <6> and prototypes <7>.
The lookup found two matching system slots which means the message is ambiguous. The error message also says where the matching slots were found.

8.8. Programmer defined errors

These are explicitly raised in the Self program to report errors, e.g. sending the message first to an empty list will cause such an error.

Error: first is absent.
Receiver is: list <7>.

Use the selectors error: and error:Arguments: to raise a programmer defined error.

8.9. Primitive errors

Primitive failures occur when a primitive cannot perform the requested operation, for example, because of a missing or invalid argument.

badTypeError: the ’_IntAdd:’ primitive failed.
Its receiver was shell <6>.
The primitive failed with badTypeError because the shell in not an integer.
The selector 12 could not be sent to shell because it is not a string.
The primitive _Perform expects a string as its first argument.
The selector ’add:’ could not be sent to shell <0> because it does not take 2 arguments.
The primitive _Perform received the wrong number of arguments.

There are many other kinds of possible primitive errors.

8.10. Nonrecoverable process errors

Errors that stop a process from continuing execution are referred to as nonrecoverable errors.

The stack has grown too big.
(Self 4.0 limits stack sizes, and cannot resume processes with stack
A stack overflow error occurs because the current version of Self allocates a fixed size stack for each process, and the stack cannot be expanded.
Self 4.0 cannot run a block after its enclosing method has returned.
(Self cannot resume this process, either.)
This error occurs if a block is executed after its lexically enclosing method has returned. This is called a “non-LIFO” block. Non-LIFO blocks are not supported by the current version of Self.

8.11. Fatal errors

In rare cases, the virtual machine may encounter a fatal error (e.g., a resource limit is exceeded or an internal error is discovered). When this happens, a short menu is displayed:

VM Version: 4.0.5, Tue 27 Jun 95 13:35:49 Solaris 2.x (svr4)
Internal error: signal 11 code 3 addr 4 pc 0x1ac768.
Do you want to:
1) Quit Self (optionally attempting to write a snapshot)
2) Try to print the Self stack
3) Try to return to the Self prompt
4) Force a core dump
Your choice:

The first two lines help the Self implementors locate the problem. Printing the Self stack may provide more information about the problem but does not always work. Returning to the Self prompt may be successful, but the system integrity may have been compromised as a result of the error. The safest course is to attempt to write a snapshot (if there are unsaved changes), and then check the integrity of the snapshot by executing the primitive _Verify after starting it. If there are any error messages from the primitive, do not attempt to continue using the snapshot.

Since fatal errors usually arise from a bug in the virtual machine, please send the Self group a bug report, and include a copy of the error message if possible. If the error is reproducible please describe how to reproduce it (including a snapshot or source files may be helpful).

8.12. The initial Self world

The diagram on the following pages shows all objects in the “bare” Self world. In addition, literals like integers, floats, and strings are conceptually part of the initial Self world; block and object literals are created by the programmer as needed. All the objects in the system are created by adding slots to these objects or by cloning them. Table 8.1 lists all the initial objects and provides a short description for each. Reading in the world rearranges the structure of the “bare” Self world (see The Self World).


Fig. 8.1 The initial Self world (part 1)


Fig. 8.2 The initial Self world (part 2)

Table 8.1 Objects in the initial Self world
Object Description
lobby The center of the Self object hierarchy, and the context in which expressions typed in at the VM prompt, read in via _RunScript, or used as the initializers of slots, are evaluated.
Table 8.2 Objects in the lobby
Object Description
shell After reading in the world, shell is the context in which expressions typed in at the prompt are evaluated.
snapshotAction An object with slot for the startup action (see System-triggered messages), postRead. This slot initially contains nil.
systemObjects This object contains slots containing the general system objects, including nil, true, false, and the prototypical vectors and mirrors.
Table 8.3 Objects in systemObjects
Object Description
nil The initializer for slots that are not explicitly initialized. Indicates “not a useful object.”
true Boolean true. Argument to and returned by some primitives.
false Boolean false. Argument to and returned by some primitives.
vector The prototype for (normal) vectors.
byteVector The prototype for byte vectors.
proxy The prototype for proxy objects.
fctProxy The prototype for fctProxy objects.
vector parent The object that vector inherits from. Since all object vectors will inherit from this object (because they are cloned from vector), this object will be the repository for shared behavior (a traits object) for vectors.
byteVector parent Similar to vector parent: the byteVector traits object.
slotAnnotation The default slot annotation object.
objectAnnotation The default object annotation object
profiler The prototype for profilers.
mirrors See below.
Table 8.4 Literals and their parents
Literal Description
integers Integers have one slot, a parent slot called parent. All integers have the same parent: see 0 parent, below.
0 parent All integers share this parent, the integer traits object.
floats Floats have one slot, a parent slot called parent. All floats have the same parent: see 0.0 parent, below.
0.0 parent All floats share this parent, the float traits object.
canonical strings In addition to a byte vector part, a canonical string has one slot, parent, a parent slot containing the same object for all canonical strings (see ’’ parent below).
'' parent All canonical strings share this parent, the string traits object.
blocks Blocks have two slots: parent, a parent slot containing the same object for all blocks (see [ ] parent, below), and value (or value:, or value:With:, etc., depending on the number of arguments the block takes) which contains the block’s deferred method.
[ ] parent All blocks share this parent, the block traits object.
Table 8.5 Prototypical mirrors
Mirror Description
smiMirror Prototypical mirror on a small integer; the reflectee is 0.
floatMirror Prototypical mirror on a float; the reflectee is 0.0.
stringMirror Prototypical mirror on a canonical string; the reflectee is the empty canonical string (’’).
processMirror Prototypical mirror on a process; the reflectee is the initial process.
byteVectorMirror Prototypical mirror on a byte vector; the reflectee is the prototypical byte vector.
objVectorMirror Prototypical mirror on object vectors; the reflectee is the prototypical object vector.
assignmentMirror Mirror on the assignment primitive; the actual reflectee is an empty object.
mirrorMirror Prototypical mirror on a mirror; the reflectee is slotsMirror.
slotsMirror Prototypical mirror on a plain object without code; the reflectee is an empty object.
blockMirror Prototypical mirror on a block.
methodMirror Prototypical mirror on a normal method.
blockMethodMirror Prototypical mirror on a block method.
methodActivationMirror Prototypical mirror on a method activation.
blockMethodActivationMirror Prototypical mirror on a block activation.
proxyMirror Prototypical mirror on a proxy.
fctProxyMirror Prototypical mirror on a fctProxy.
profilerMirror Prototypical mirror on a profiler.

All of the prototypical mirrors consist of one slot, a parent slot named parent. Each of these parent slots points to an empty object (denoted in Fig. 8.1 by “( )”).

8.13. Option Primitives

This section has not been updated to include all options present in Self 4.0.

Option primitives control various aspects of the Self system and its inner workings. Many of them are used to debug or instrument the Self system and are probably of little interest to users. The options most useful for users are listed in Table 8.6; other option primitives can be found in Appendix 10.8 Primitives, and a list of all option primitives and their current settings can be printed with the primitive _PrintOptionPrimitives.

Table 8.6 Some useful option primitives
Name Description
_PrintPeriod[:] Print a period when reading a script file with _RunScript. Default: false.
_PrintScriptName[:] Print the file name when reading a script file. Default: false.
_Spy[:] Start the system monitor (see Appendix 10.7 for details). Default: false.
_StackPrintLimit[:] Controls the number of stack frames printed by _PrintProcessStack. Default: 20.
_DirPath[:] The default directory path for script files.

Each option primitive controls a variable within the virtual machine containing a boolean, integer, or string (in fact, the option primitives can be thought of as “primitive variables”). Invoking the version of the primitive that doesn’t take an argument 1 returns the current setting; invoking it with an argument sets the variable to the new value and returns the old value.

Try running the system monitor with _Spy: true. The system monitor will continuously display various information about the system’s activities and your memory usage.

8.14. Interfacing with other languages

This chapter describes how to access objects and call routines that are written in other languages than Self. We will refer to such entities as foreign objects and foreign routines. A typical use would be to make a function found in a C library accessible in Self. Three steps are necessary to accomplish this:

  • Write and compile a piece of “glue” code that specifies argument and result types for the foreign routine and how to convert between these types and Self objects.
  • Link the resulting object code to the Self virtual machine.
  • Create a function proxy object (actually a foreignFct object) that represents the routine in the Self world.

Each of these steps is described in detail in the following sections.

8.14.1. proxy and fctProxy objects

A foreign object is represented by a proxy object in the Self world. A proxy object is an object that encapsulates a pointer to the foreign object it represents. In addition to the pointer to the foreign object, the proxy object contains a type seal. A type seal is an immutable value that is assigned to the proxy object, when it is created. The type seal is intended to capture type information about the pointer encapsulated in the proxy. For example, proxies representing window objects should have a different type seal than proxies representing event objects. By checking the type seal against an expected value whenever a proxy is “opened”, many type errors can be caught. The last property of proxy objects is that they can be dead or live. If an attempt is made to use the pointer in a dead proxy object, an error results (deadProxyError). Proxy objects may be explicitly killed, by sending the primitive message _Kill to them. Furthermore, they are automatically killed after reading in a snapshot. This way problems with dangling references to foreign objects that were not included in the snapshot are avoided.

fctProxy objects are similar to proxy objects: they have a type seal and are either live or dead. However, they represent a foreign routine, rather than a foreign object. A foreign routine can be invoked by sending the primitive messages _Call, _Call:{With:}, _CallAndConvert{With:And:} to the fctProxy representing it. Note that fctProxy objects are low-level. Most, if not all, uses of foreign routines should use the interface provided by foreignFct objects.

Proxies (and fctProxies) can be freely cloned. However a cloned proxy will be dead. A dead proxy is revived when it is used by a foreign function to, e.g., return a pointer. The return value of the foreign function together with a type seal is stored into the dead proxy, which is then revived and returned as the result of the foreign routine call. The motivation for this somewhat complicated approach is that there will be several different kinds of proxies in a typical Self system. Different kinds of proxies may have different slots added, so rather than having the foreign routine figure out which kind of proxy to clone for the result, the Self code calling the foreign routine must construct and pass down an “empty” (dead) proxy to hold the result. This proxy is called a result proxy and it is the last argument supplied to the foreign function.

8.14.2. Glue code

Glue code is responsible for the transition from Self to foreign routines. It forms wrappers around foreign routines. There is one wrapper per foreign routine. A wrapper takes a number of arguments of type oop, and returns an oop (oop is the C++ type for “reference to Self object”). When a wrapper is executed, it performs the following steps:

  1. Check that the arguments supplied have the correct types.
  2. Convert the arguments from Self representation to the representation that the foreign routine needs.
  3. Invoke the foreign routine on the converted arguments.
  4. Convert the return value of the foreign routine to a Self object and return this as the Self level result.

To make it easier to write glue code, a special purpose language has been designed for this. The result is that glue for a foreign routine will often consist of only a single line. The glue language is implemented as a set of C++ preprocessor macros. Therefore, glue code is just a (rather peculiar) kind of C++. Glue code can be in a file of its own, or – if it is glue for calling C++ routines – it can be in the same file as the foreign routines, and compiled with them.

To make the definition of the glue language available, the file containing glue code must contain:

# include "_glueDefs.c.incl"

The file “_glueDefs.c.incl” includes a bunch of C++ header files that contain all the definitions necessary for the glue. Of the included files, “glueDefs.h” is probably the most interesting in this context. It defines the glue language and also contains some comments explaining it.

Since different foreign languages have different type systems and calling conventions the glue language is actually not a single language, but one for each supported foreign language. Presently C and C++ are supported. See sections C glue and C++ glue for details.

8.14.3. Compiling and linking glue code

Since glue code is a special form of C++ code, a C++ compiler is needed to translate it. The way this is done may depend on the computer system and the available C++ compiler. The following description applies to Sun SPARCstations using the GNU g++ compiler.

A specific example of how to compile glue code can be found in the directory containing the toself demo (see A complete application using foreign functions for further details). The makefile in that directory describes how to translate a .c file containing glue into something that can be invoked from Self. This is a two stage process: first the .c file is compiled into a .o file which is then linked (perhaps with other .o files and libraries that the glue code depends on) into a .so file (a so-called dynamic library). While the compilation is straightforward, several issues concerning the linking must be explained.

Before a foreign routine can be called it must be linked to the Self virtual machine. The linking can be done either statically, i.e. before Self is started, or dynamically, i.e. while Self is running. The Self system employs both dynamic and static linking, but users should only use dynamic linking, as static linking requires more understanding of the structure of the Virtual Machine. The choice between dynamic and static linking involves a trade-off between safety and flexibility as outlined in the following.
Dynamic linking
Dynamic linking has the advantage that it is done on demand, so only foreign routines that are actually used in a particular session will be loaded and take up space. Debugging foreign routines is also easier, especially if the dynamic linker supports unlinking. The main disadvantages with dynamic linking is that more things can go wrong at run time. For example, if an object file containing a foreign routine can not be found, a run time error occurs. The Sun OS dynamic linker, ld.so, only handles dynamic libraries which explains why the second stage of glue translation is necessary.
Static linking
Static linking, the alternative that was not chosen for Self, has the advantage that it needs to be done only once. The statically linked-in files will then be available for ever after. The main disadvantages are that the linked-in files will always take up space whether used or not in a given Self session, that the VM must be completely relinked every time new code is added, and that debugging is harder because there is no way to unlink code with bugs in. For these reasons the following examples all use dynamic linking.

8.14.4. A simple glue example: calling a C function

Suppose we have a C function that encrypts text strings in some fancy way. It takes two arguments, a string to encrypt and a key, and returns a string which is the result of the encryption. To use this function from Self, we write a line of C glue. Here is the entire file, “encrypt.c”, containing both the encryption function and the glue:

/* Make glue available by including it. */
# include "incls/_glueDefs.c.incl"
/* Naive encryption function. */
char *encrypt(char *str, int key) {
    static char res[1000];
    int i;
    for (i = 0; str[i]; ++i)
        res[i] = str[i] + key;
    res[i] = ’\0’;
    return res;

/* Make glue expand to full functions, not just prototypes. */
    C_func_2(string,, encrypt, encrypt_glue,, string,, int,)
# undef WHAT_GLUE

A few words of explanation: the last three lines of this file contain the glue code. First defining WHAT_GLUE to be FUNCTIONS, makes the following line expand into a full wrapper function (defining WHAT_GLUE to be PROTOTYPES instead, will cause the C_func_2 line to produce a function prototype only). The line containing the macro C_func_2 is the actual wrapper for encrypt. The “2” designates that encrypt takes 2 arguments. The meaning of the arguments, from left to right are:

  • string,: specifies that encrypt returns a string argument.
  • encrypt: name of function we are constructing wrapper for.
  • encrypt_glue: name that we want the wrapper function to have.
  • An empty argument signifying that encrypt is not to be passed a failure handle (explained later).
  • string,: specifies that the first argument to encrypt is a string.
  • int,: specifies that the second argument to encrypt is an int.

Having written this file, we now prepare a makefile to compile and link it. To do this, we can extend the makefile in objects/glue/{sun4,svr4} (depending on OS in use) and then run make. This results in the shared library file encrypt.so. Finally, to try it out, we can type these commands (at the Self prompt or in the UI):

> _AddSlotsIfAbsent: ( | encrypt | )

> encrypt: ( foreignFct copyName: ’encrypt_glue’ Path: ’encrypt.so’ )

> encrypt
<C++ function(encrypt_glue)>

> encrypt value: ’Hello Self’ With: 3

> encrypt value: ’Khoor#Vhoi’ With: -3
’Hello Self’

Comparing the signature for the function encrypt with the arguments to the C_func_2 macro it is clear that there is a straightforward mapping between the two. One day we hope to find the time to write a Self program that can parse a C or C++ header file and generate glue code corresponding to the definitions in it. In the meantime, glue code must be handwritten.

8.14.5. C glue

C glue supports accessing C functions and data from Self. There are three main parts of C glue:

  • Calling functions.
  • Reading/assigning global variables.
  • Reading/assigning a component in a struct that is represented by a proxy object in Self.

In addition, C++ glue for creating objects can be used to create C structs (see section C++ glue). The following sections describe each of these parts of C glue.

8.14.6. Calling C functions

The macro C_func_N where N is 0, 1, 2, ... is used to “glue in” a C function. The number N denotes the number of arguments that should be given at the Self level, when calling the function. This number may be different from the number of arguments that the C function takes since, e.g., some argument conversions (see below) produce two C arguments from one Self object. Here is the general syntax for C_func_N:

C_func_N(res_cnv,res_aux, fexp, gfname, fail_opt, c0,a0, ... cN,aN)

Compare this with the glue that was used in the encrypt example in section A simple glue example: calling a C function:

C_func_2(string,, encrypt, encrypt_glue,, string,, int,)

The meaning of each argument to C_func_N is as follows:

  • res_cnv,res_aux: these two arguments form a “conversion pair” that specifies how the result that the function returns is converted to a Self object. In the encrypt example, where the function returns a null terminated string, res_cnv has the value string, and res_aux is empty. Table 8.7 lists all the possible values for the res_cnv,res_aux pair.
  • fexp is a C expression which evaluates to the function that is being glued in. In the simplest case, such as in the encrypt example, the expression is the name of a function, but in general it may be any C expression, involving function pointers etc., which in a global context evaluates to a function.
  • gfname: the name of the function which the C_func_N macro expands into. In the encrypt example, the convention of appending _glue to the C function’s name was used. When accessing a glued-in function from Self, the value of gfname is the name that must be used.
  • fail_opt: there are two possible values for this argument. It can be empty (as in the example) or it can be fail. In the latter case, the C function being called is passed an additional argument that will be the last argument and have type “void *”. Using this argument, the C function may abort its execution and raise an exception. The result is that the “IfFail block” in Self will be invoked.
  • ci,ai: each of these pairs describes how to convert a Self level argument to one or more C level arguments. For example, in the glue for encrypt, c0,``a0`` specifies that the first argument to encrypt is a string. Likewise c1,``a1`` specifies that the second argument is an integer. Note that in both these cases, the a-part of the conversion is empty. Table 8.7 lists all the possible values for the ci,``ai`` pair.

Handling failures. Here is a slight modification of the encryption example to illustrate how the C function can raise an exception that causes the “IfFail block” to be invoked at the Self level:

/* Make glue available by including it. */
# include "incls/_glueDefs.c.incl"
/* Naive encryption function. */
char *encrypt(char *str, int key, void *FH) {
    static char res[1000];
    int i;
    if (key == 0) {
            failure(FH, "key == 0 is identity map");
            return NULL;
    for (i = 0; str[i]; i++)
            res[i] = str[i] + key;
    res[i] = ’\0’;
    return res;
/* Make glue expand to full functions, not just prototypes. */
    C_func_2(string,, encrypt, encrypt_glue, fail, string,, int,)
# undef WHAT_GLUE

Observe that the fail_opt argument now has the value fail and that the encrypt function raises an exception, using failure, if the key is 0. There are two ways to raise exceptions:

extern "C" void failure(void *FH, char *msg);
extern "C" void unix_failure(void *FH, int err = -1);

In both cases, the FH argument is the “failure handle” that was passed by the C_func_N macro. The second argument to failure is a string. It will be passed to the “IfFail block” in Self. unix_failure takes an optional integer as its second argument. If this integer has the value -1, or is missing, the value of errno is used instead. The integer is interpreted as a UNIX error number, from which a corresponding string is constructed. The string is then, as for failure, passed to the “IfFail block” at the call site in Self.


After calling failure or unix_failure a normal return must be done. The value returned (in the example NULL) is ignored.

8.14.7. Reading and assigning global variables

Reading the value of a global variable is done using the C_get_var macro. Assigning a value to a global variable is done using C_set_var. Both macros expand into a C++ function that converts between Self and C representation, and reads or assigns the variable. Here is the general syntax:

C_get_var(cnvt_res,aux_res, expr, gfname)
C_set_var(var, expr_c0,expr_a0, gfname)

A concrete example is reading the value of the variable errno, which can be done using:

C_get_var(int,, errno, get_errno_glue)

The meaning of the each argument is:

  • cnvt_res,``aux_res``: how to convert the value of the global variable that is being read to a Self object. In the errno example, cnvt_res is int and aux_res is empty, since the type of errno is int. The cnvt_res,``aux_res`` can be any one of the result conversions found in Table 8.7.
  • expr is the variable whose value is being read. In the errno example, it is simply errno, but in general, it may actually be any expression that is valid in a global context, even an expression involving function calls.
  • gfname: the name of the C++ function that C_get_var or C_set_var expands into.
  • var is the name of a global variable that a value is assigned to. In general, var, may be any expression that in a global context evaluates to an l-value.
  • expr_c0,``expr_a0``: when assigning to a variable, the value it is assigned is obtained by converting a Self object to a C value. The expr_c0,``expr_a0`` pair, which can be any one of the argument conversions listed in Table 8.7, specifies how to do this conversion.

8.14.8. Reading and assigning struct components

Reading the value of a struct component or assigning a value to it is similar to doing the same operations on a global variable. The difference is that the struct must somehow be specified. This is taken care of by the macros C_get_comp and C_set_comp. The general syntax is:

C_get_comp(cnvt_res,aux_res, cnvt_strc,aux_strc, comp, gfname)
C_set_comp(cnvt_strc,aux_strc, comp, expr_c0,expr_a0, gfname)

Here is an example, assigning to the sin_port field of a struct sockaddr_in (this struct is defined in /usr/include/netinet/in.h):

struct sockaddr_in {
    short           sin_family;
    u_short         sin_port;
    struct in_addr  sin_addr;
    char            sin_zero[8];

The struct is represented by a proxy object:

char *socks = "type seal for sockaddr_in proxies";
C_set_comp(proxy,(sockaddr_in *,socks), .sin_port, short,,set_sin_port_glue)

The sockaddr_in example defines a function, set_sin_port_glue, which can be called from Self. The function takes two arguments, the first being a proxy representing a sockaddr_in struct, the second being a short integer. After converting types, set_sin_port_glue performs the assignment:

(*first_converted_arg).sin_port = second_converted_arg.

In general the meaning of the C_get_comp and C_set_comp arguments is:

  • cnvt_res, aux_res: how to convert the value of the component that is being read to a Self object. Any of the result conversions found in Table 8.7 may be applied.
  • cnvt_strc, aux_strc: the conversion that is applied to produce a struct upon which the operation is performed. In the sin_port example, this conversion is a proxy conversion, implying that in Self, the struct whose sin_port component is assigned is represented by a proxy object. In general, any of the argument conversions from Table 8.7 that results in a pointer, may be used.
  • comp is the name of the component to be read or assigned. In the sin_port example, this name is “.sin_port”. Note that it includes a “.”. This, e.g., allows handling pointers to int’s by pretending that it is a pointer to a struct and operating on a component with an empty name.
  • gfname: the name of the C++ function that C_get_comp or C_set_comp expands into.
  • expr_co, expr_a0: when assigning to a component, the value it is assigned is obtained by converting a Self object to a C value. The expr_co, expr_a0 pair, which can be any one of the argument conversions listed in Table 8.7, specifies how to do this conversion.

8.14.9. C++ glue

Since C++ is a superset of C, all of C glue can be used with C++. In addition, C++ glue provides support for:

  • Constructing objects using the new operator.
  • Deleting objects using the delete operator.
  • Calling member functions on objects.

Each of these parts will be explained in the following sections.

8.14.10. Constructing objects

In C++, objects are constructed using the new operator. Constructors may take arguments. The macros CC_new_N where N is a small integer, support calling constructors with or without arguments. Calling a constructor is similar to calling a function, so for additional explanation, please refer to section Calling C functions. Here is the general syntax for constructing objects using C++ glue:

CC_new_N(cnvt_res,aux_res, class, gfname, c0,a0, c1,a1, ... cN,aN)

For example, to construct a sockaddr_in object, the following glue statement could be used:

CC_new_0(proxy,(sockaddr_in *,socks), sockaddr_in, new_sockaddr_in)

The meanings of the CC_new_N arguments are as follows:

  • cnvt_res, aux_res: the result of calling the constructor is an object pointer. The result conversion pair cnvt_res, aux_res (see Table 8.7), specifies how this pointer is converted to a Self object before being returned. In the sockaddr example, the proxy result conversion is used.
  • class is the name of the class (or struct) that is being instantiated.
  • gfname: the name of the C++ function that the CC_new_N macro expands into.
  • ci, ai: if the constructor takes arguments, these arguments must be converted from Self representation to C++ representation. The arguments conversion pairs ci, ai specify how each argument is converted. See Table 8.7 for a description of all argument conversions. In the sockaddr example, there are no arguments.

8.14.11. Deleting objects

C++ objects can have destructors that are executed when the objects are deleted. To ensure that the destructor is called properly, the delete operator must know the type of the object being deleted. This is ensured by using the CC_delete macro, which has the following form:

CC_delete(cnvt_obj,aux_obj, gfname)

For example, to delete sockaddr_in objects (constructed as in the previous section), the CC_delete macro should be used in this manner:

CC_delete(proxy,(sockaddr_in *,socks), delete_sockaddr_in)

In general, the meaning of the arguments given to CC_delete is:

  • cnvt_obj,aux_obj: this pair can be any of the argument conversions found in Table 8.7 that produces a pointer to the object that will be deleted.
  • gfname: the name of the C++ function that this invocation of CC_delete expands into.

8.14.12. Calling member functions

Table 8.7 lists all the available argument conversions. Each row represents one conversion, with the first two columns designating the conversion pair. The third column lists the types of Self objects that the conversion pair accepts. The fourth column lists the C types that it produces. The fifth column lists the kind of errors that can occur during the conversion. Finally, the sixth column contains references to numbered notes. The notes are found in the paragraphs following the table.

Calling member functions is similar to calling “plain” functions, so please also refer to section Calling C functions. The difference is that an additional object must be specified: the object upon which the member function is invoked (the receiver in Self terms). Calling a member function is accomplished using one of the macros:

CC_mber_N(cnvt_res,aux_res, cnvt_rec,aux_rec, mname, gfname,
          fail_opt, c0,a0, c1,a1, ..., cN,aN)

For example here is how to call the member function zock on a sockaddr_in object given by a proxy:

CC_mber_0(bool,, proxy,(sockaddr_in *,socks), zock, zock_glue,)

The arguments to CC_mber_N are:

  • cnvt_res, aux_res: this pair, which can be any of the result conversions from Table 8.7, specifies how to convert the result of the member function before returning it to Self. For example, the zock member function returns a boolean.
  • cnvt_rec, aux_rec: the object on which the member function is invoked. Often this will be a proxy conversion as in the zock example.
  • mname is the name of the member function. In general, it may be any expression, such that receiver->mname evaluates to a function.
  • gfname is the name of the C++ function that the CC_mber_N macro expands into.
  • fail_opt: whether or not to pass a failure handle to the member function (refer to section Calling C functions for details).
  • ci, ai: these are argument conversion pairs specifying how to obtain the arguments for the member function. Any conversion pair found in Table 8.7 may be used.

8.14.13. Conversion pairs

A major function of glue code is to convert between Self objects and C/C++ values. This conversion is guarded by so-called conversion pairs. A conversion pair is a pair of arguments given to a glue macro. It handles converting one or at most a few types of objects/values. There are different conversion pairs for converting from Self objects to C/C++ values (called argument conversion pairs) and for converting from C/C++ values to Self objects (called result conversion pairs).

8.14.14. Argument conversions – from Self to C/C++

An argument conversion is given a Self object and performs these actions to produce a corresponding C or C++ value:

  • check that the Self object it has been given is among the allowed types. If not, report badTypeError (invoke the failure block (if present) with the argument ’badTypeError’).
  • check that the object can be converted to a C/C++ value without overflow or any other error. If not, report the relevant error.
  • do the conversion, i.e., construct the C/C++ value corresponding to the given Self object.
Table 8.7 Argument conversions - from Self to C/C++
Conversion Second part Self type C/C++ type Errors Notes
bool   boolean int (0 or 1) badTypeError  
char   smallInt char badTypeError overflowError 1
signed_char   smallInt signed char badTypeError overflowError  
unsigned_char   smallInt unsigned char badSignError badTypeError overflowError  
short   smallInt short badTypeError overflowError  
signed_short   smallInt signed short badTypeError overflowError  
unsigned_short   smallInt unsigned short badSignError badTypeError overflowError  
int   smallInt int badTypeError  
signed_int   smallInt signed int badTypeError  
unsigned_int   smallInt unsigned int badSignError badTypeError  
long   smallInt long badTypeError  
signed_long   smallInt signed long badTypeError  
unsigned_long   smallInt unsigned long badSignError  
smi   smallInt smi badTypeError 2
unsigned_smi   smallInt smi badSignError badTypeError 2
Conversion Second part Self type C/C++ type Errors Notes
float   float float badTypeError 3
double   float double badTypeError 3
long_double   float long double badTypeError 3
bv ptr_type byte vector ptr_type badTypeError 4
bv_len ptr_type byte vector ptr_type, int badSizeError badTypeError 4, 5
bv_null ptr_type byte vector/0 ptr_type badTypeError 4, 6
bv_len_null ptr_type byte vector/0 ptr_type, int badSizeError badTypeError 4, 5, 6
cbv ptr_type byte vector ptr_type badTypeError 7
cbv_len ptr_type byte vector ptr_type, int badSizeError badTypeError 7
cbv_null ptr_type byte vector/0 ptr_type badTypeError 7
cbv_len_null ptr_type byte vector/0 ptr_type, int badSizeError badTypeError 7
string   byte vector char * badTypeError nullCharError 8
string_len   byte vector char *, int badTypeError nullCharError 5, 8
string_null   byte vector/0 char * badTypeError nullCharError 6, 8
string_len_null   byte vector/0 char *, int badTypeError nullCharError 5, 6, 8
proxy (ptr_type, type_seal) proxy ptr_type, != NULL badTypeError badTypeSealError, deadProxyError,nullPointerError 9
proxy_null (ptr_type, type_seal) proxy ptr_type badTypeError badTypeSealError deadProxyError 9
any_oop   any object oop   10
oop oop subtype corresponding object oop (subtype) badTypeError 11
any C/C++ type int/float/proxy/byte-vector, int int/float/ptr/ptr badIndexError badTypeError deadProxyError 12 Notes

  1. The C type char has a system dependent range. Either 0..255 or -128..127.
  2. The type smi is used internally in the virtual machine (a 30 bit integer).
  3. Precision may be lost in the conversion.
  4. The second part of the conversion is a C pointer type. The address of the first byte in the byte vector, cast to this pointer type, is passed to the foreign routine. It is the responsibility of the foreign routine not to go past the end of the byte vector. The foreign routine should not retain pointers into the byte vector after the call has terminated. Note: canonical strings can not be passed through a bv conversion (badTypeError will result). This is to ensure that they are not accidentally modified by a foreign function.
  5. This conversion passes two values to the foreign routine: a pointer to the first byte in the byte vector, and an integer which is the length of the byte vector divided by sizeof(*ptr_type). If the size of the byte vector is not a multiple of sizeof(*ptr_type), badSizeError results.
  6. In addition to accepting a byte vector, this conversion accepts the integer 0, in which case a NULL pointer is passed to the foreign routine.
  7. The cbv conversions are like the bv conversions except that canonical strings are allowed as actual arguments. A cbv conversion should only be used if it is guaranteed that the foreign routine does not modify the bytes it gets a pointer to.
  8. All the string conversions take an incoming byte vector, copy the bytes part, add a trailing null char, and pass a pointer to this copy to the foreign routine. After the call has terminated, the copy is discarded. If the byte vector contains a null char, nullCharError results.
  9. The type_seal is an int or char * expression that is tested against the type seal value in the proxy. If the two are different, badTypeSealError results. The special value ANY_SEAL will match the type seal in any proxy. Note that the proxy conversion will fail with nullPointerError if the proxy object it is given encapsulates a NULL pointer.
  10. The any_oop conversion is an escape: it passes the Self object unchanged to the foreign routine.
  11. The oop conversion is mainly intended for internal use. The second argument is the name of an oop subtype. After checking that the incoming argument points to an instance of the subtype, the pointer is cast to the subtype.
  12. The any conversion is different from all other conversions in that it expects two incoming Self objects. The actions of the conversion depends on the type of the first object in the following way. If the first object is an integer, the second argument must also be an integer; the two integers are converted to C int’s, the second is shifted 16 bits to the left and they are or’ed together to produce the result. If the first object is a float, it is converted to a C float and the second object is ignored. If the first object is a proxy, the result is the pointer represented by the proxy, and the second argument is ignored. If the first object is a byte vector, the second object must be an integer which is interpreted as an index into the byte vector; the result is a pointer to the indexed byte.

8.14.15. Result conversions - from C/C++ to Self

A result conversion is given a C or C++ value of a certain type and performs these actions to produce a corresponding Self object:

  • check that the C/C++ value can be converted to a Self object with no overflow or other error occurring. If not, report the error.
  • do the conversion, i.e., construct the Self object corresponding to the given C/C++ value.

Table 8.8 lists all the available result conversions. Each row represents one conversion, with the first two columns designating the conversion pair. The third column lists the type of C or C++ value that the conversion pair accepts. The fourth column lists the type of Self object the conversion produces. The fifth column lists the kind of errors that can occur during the conversion. Finally, the sixth column contains references to numbered notes. The notes are found in the paragraphs following the table.

Table 8.8 Result conversions - from C/C++ to Self
Conversion Second part C/C++ type Self type Errors Notes
void   void smallInt (0)    
bool   int boolean    
char   char smallInt    
signed_char   signed char smallInt    
unsigned_char   unsigned char smallInt    
short   short smallInt    
signed_short   signed short smallInt    
unsigned_short   unsigned short smallInt    
int   int smallInt overflowError  
signed_int   signed int smallInt overflowError  
unsigned_int   unsigned int smallInt overflowError  
long   long smallInt overflowError  
signed_long   signed long smallInt overflowError  
unsigned_long   unsigned long smallInt overflowError  
smi   smi smallInt overflowError  
int_or_errno n int int a UNIX error 1
float   float float   2
double   double float   2
long_double   long double float   2
string   char * byte vector nullPointerError 3
proxy (ptr_type, type_seal) ptr_type proxy nullPointerError 3, 4, 8
proxy_null (ptr_type, type_seal) ptr_type proxy   4, 8
proxy_or_errno (ptr_type, type_seal, n) ptr_type proxy a UNIX error 4, 5, 8
fct_proxy (ptr_type, type_seal, arg_count) ptr_type fctProxy nullPointerError 3, 6, 8
fct_proxy_null (ptr_type, type_seal, arg_count) ptr_type fctProxy   6, 8
oop   oop corresponding object   7, 8 Notes

  1. This conversion returns an integer value, unless the integer has the value n (the second part of the conversion; often -1). If the integer is n, the conversion interprets the return value as a UNIX error indicator. It then constructs a string describing the error (by looking at errno) and invokes the “IfFail block” with this string.
  2. Precision may be lost.
  3. This conversion fails with nullPointerError if attempting to convert a NULL pointer.
  4. The ptr_type is the C/C++ type of the pointer. The type_seal is an expression of type int or char *.The conversion constructs a new proxy object, stores the C/C++ pointer in it and sets its type seal to be the value of type_seal.
  5. If the pointer is n (often n is NULL), the conversion fails with a UNIX error, similar to the way int_or_errno may fail.
  6. The fct_proxy, fct_proxy_null and fct_proxy_or_errno conversions are similar to the corresponding proxy conversions. The difference is that they produce a fctProxy object rather than a proxy object. Also, their second part is a triple rather than a pair. The extra component specifies how many arguments the function takes, if called. The special keyword unknownNoOfArgs or any nonnegative integer expression can be used here.
  7. This conversion is an escape: it passes the C value unchanged to Self. It is an error to use it if the C value is not an oop.
  8. The proxy (fctProxy) object that is returned by these conversions is not being created by the glue code. Rather a proxy (fctProxy) must be passed down from the Self level. This proxy (fctProxy), a result proxy, will then be side effected by the glue: the value that the foreign function returns will be stored in the result proxy together with the requested type seal. It is required that the result proxy is dead when passed down (else a liveProxyError results). After being side-effected and returned, the result proxy is live. The result proxy is the last argument of the function that the glue macro expands to.

8.14.16. A complete application using foreign functions

This section gives a description of a complete application which uses foreign functions. The aim is to present a realistic and complete example of how foreign functions may be used. The complete source for the example is found in the directory objects/applications/serverDemo in the Self distribution.

The example used is an application that allows Self expressions to be easily evaluated by non- Self processes. Having this, it then becomes possible to start Self processes from a UNIX prompt (shell) or to specify pipe lines in which some of the processes are Self processes. For example in

proto% cat someFile | tokenize | sort -r | capitalize | tee lst

it may be the case that the filters tokenize and capitalize perform most of their work in Self. Likewise, the command

proto% mail

may invoke some fancy mail reader written in Self rather than the standard UNIX mail reader.

To see how the above can be accomplished, please refer to Fig. 8.3 below. The left side of the figure shows the external view of a typical UNIX process. It has two files: stdin and stdout (for simplicity we ignore stderr). Stdin is often connected to the keyboard so that characters typed here can be read from the file stdin. Likewise, stdout is typically connected to the console so that the process can display output by writing it to the file stdout. Stdin and stdout can also be connected to “regular” files, if the process was started with redirection. The right side of Fig. 8.3 shows a two stage pipe line. Here stdout of the first process is connected to stdin of the second process.


Fig. 8.3 A single UNIX process and an pipe line.

Fig. 8.3 illustrates a simple trick that in many situations allows Self processes to behave as if they are full-fledged UNIX processes. A Self process is represented by a “real” UNIX process which transparently communicates with the Self process over a pair of connected sockets. The communication is bidirectional: input to the UNIX process is relayed to the Self process over the socket connection, and output produced by the Self process is sent over the same socket connection to the UNIX process which relays it to stdout. The right part of Fig. 8.3 shows how the UNIX/Self process pair can fit seamlessly into a pipe line.


Fig. 8.4 A Self process and how it fits into a pipe line.

Source code that facilitates setting up such UNIX/Self process pairs is included in the Self distribution. The source consists of two parts: one being a Self program (called server), the other being a C++ program (called toself). When the server is started, it creates a socket, binds a name to it and then listens for connections on it. toself establishes connections to the server program. The first line that is transmitted when a connection has been set up goes from toself to the server. The line contains a Self expression. Upon receiving it, the server forks a new process to evaluate the expression in the context of the lobby augmented with a slot, stdio, that contains a unixFile-like object that represents the socket connection. When the forked process terminates, the socket connection is shut down. The toself UNIX process then terminates.

The Self expression that forms the Self process is specified on the command line when toself is started. For example, if the server has been started, the following can be typed at the UNIX prompt:

proto% toself stdio writeLine: 5 factorial printString

proto% echo something | toself capitalize: stdio

proto% toself capitalize: stdio
Write some text that goes to stdin of the toself program
More text


If you want to try out these examples, locate the files server.self, socks.so and toself. The path name of the file socks.so is hardwired in the file server.self so please make sure that it has been set correctly for your system. Then file in the world and type [server start] fork at the Self prompt. Now you can go back to the UNIX prompt and try out the examples shown above.

8.14.17. Outline of toself

toself is a small C++ program found in the file toself.c. It operates in the three phases outlined above:

  1. Try to connect to a well-known port number on a given machine (the function establishConnection does this).
  2. Send the command line arguments over the connection established in 1 (the safeWrite call in main does this).
  3. While there is more input and the Self process has not shut down the socket connection, relay from stdin to the socket connection and from the socket connection to stdout (the function relay does this).

8.14.18. Outline of server

The server is a Self program. It is found in the file server.self. When the server is started, the following happens:

  1. Create a socket, bind a name to it and start listening.
  2. Loop: accept a connection and fork a new process (both step 1 and 2 are performed by the method server start). The forked process executes the method server handleRequest which:
    1. Reads a line from the connection.
    2. Sets up a context with a slot stdio referring to the connection.
    3. Evaluates the line read in step (a) in this context.
    4. Closes the connection.

8.14.19. Foreign functions and glue needed to implement server

The server program needs to do a number of UNIX calls to create sockets and bind names to them etc. The calls needed are socket, bind, listen, accept and shutdown. The first three of these are only called in a fixed sequence, so to make things easier, a small C++ function socket_bind_listen, that bundles them up in the right sequence, has been written. The accept function is more general than what is needed for this application, so a wrapper function, simple_accept, has been written. The result is that the server needs to call only three foreign functions: socket_bind_listen, simple_accept and shutdown. Glue for these three functions and the source for the first two is found in the file socks.c. This file is compiled and linked using the Makefile. The result is a shared object file, socks.so.

8.14.20. Use of foreign functions in server.self

The server program is implemented using foreignFct objects. There is only a few lines of code directly involved in setting this up. First the foreignFct prototype is cloned to obtain a “local prototype”, called socksFct, which contains the path for the socks.so file. socksFct is then cloned each time a foreignFct object for a function defined in socks.so is needed. For example, in traits socket, the following method is found:

copyPort: portNumber = ( "Create a socket, do bind, then listen."
        | sbl = socksFct copyName: ’socket_bind_listen_glue’. |
        sbl value: portNumber With: deadCopy.

This method copies a socket object and returns the copy. The local slot sbl is initialized to a foreignFct object. The body of the method simply sends value:With: to the foreignFct object. The first argument is the port number to request for the socket, the second argument is a deadCopy of self (socket objects are proxies and socket_bind_listen returns a proxy, so it must be passed a dead proxy to revive and store the result in; see section Proxy and fctProxy objects).

There are only three uses of foreignFct objects in the server and in all three cases, the foreignFct object is encapsulated in a method as illustrated above.

In general the design of foreignFct objects has been aimed at making the use of them light weight. When cloning them, it is only necessary to specify the minimal information: the name of the foreign function. They can be encapsulated in a method thus localizing the impact of redesigns. The complications of dynamic loading and linking are handled automatically, as is the recovery of dead fctProxies.


[1]The bracketed colon indicates that the argument is optional (i.e., there are two versions of the primitive, one taking an argument and one not taking an argument). The bracket is not part of the primitive name. See text for details.