revert between 56095 -> 55830 in arch

[AROS.git] / rom / oop / oop.doc

Overview:
This document tries to explain the choices that
are made in deciding the design of the new OOP
system, which will be used fo (among other things)
the HIDDs.

General design goals:

The system should..

- have fast method invocation. Especially when calling the same method
on the same objetc many times. (Eg. GetPixel()/SetPixel())


- allow non-centralized allocation of IDs. This means that there is no
agency somewhere in the world where you have to allocate an ID. IDs
are allocated at runtime.


- be location transparent. Ie. you invoke a
method the same way as for a local objet, even if
the object is owned by another task, or is
located on a different machine.


- support transparent object migration. This means that the object, the
system or the user can decide that the object should be moved or copied
to another host. Uses: If an object is accessed more often from a remote
host than locally, it can create a copy of itself. If the object is deleted
locally, it can move itself to the remote host.


- Exceptions.


- Thread safe method invocation.


Definitions:

Class - A collection of Interfaces and a description of the Objects
of this Class.

Object - Instance of a class. If you think of a Class as a blueprint of
something, then an Object is what you get when you build the something
after the instructuctions of the blueprint.

Method - A method is a function which is called with a class, an object
and a message as arguments.

Interface - An interface is a view of an object. Interfaces are
defined globally. A class can implement any number of Interfaces.
An Interface contains a table of Methods.

Method Stub - This is a normal C function which is called with an
object and a message. It is used to simplify the call of a method.
The Method Stub extracts the class from the Object and the Interface
from the Full Method ID and invokes the Method.

Method Base - The lowest ID of all methods on an Interface. Along with the
Method Count, this is used to determine if a method if supported by an
Interface. This is a constant during the lifetime of the system.

Method Count - The number of methods of an Interface. This value
can change (it is possible to add methods to an interface during
runtime).

Interface ID - The ID of an Interface. It is used to look up an
Interface when a Method should be invoked.
This is a constant during the lifetime of the system.

Method Offset - The offset to the Method Base for a Method. Always
positive. This is a constant during the lifetime of the system.

Method ID - The addition of Method Base and Method Offset for a specific
Method. This is a constant during the lifetime of the system.
If you have an Interface Object, you can call the method.

Full Method ID - The addition of Method ID and the Interface ID. The
Interface ID is used to look up the Interface and then the Method ID is
used to call the method in the Interface. This is a constant during the
lifetime of the system.

Method Name - A string with the name of a method.

Method Object - An Object used to store a Method to allow to invoke
the method fast.

Interface Object - An object used to store an Interface
which allows you to invoke all methods of that interface fast.


Possible solutions to design goals w/pros & cons

1) Invocation speed.
-------------------
For having ultimately fast method invocation we supply method objects.

Example of usage of method objects:

-------------

struct HIDDP_Gfx_SetPixel sp_msg;

struct TagItem mobj_tags[] =
{
    {A_Method_TargetObject, (IPTR)gfxhidd },
    {A_Method_Message,      (IPTR)&sp_msg},
    {A_Method_MethodID,     HIDDM_Gfx_SetPixel},
    {TAG_DONE, }
};
..

/* If you don't pass a message, then memory for one could be allocated,
and disposed in DisposeObject(spmethod).
The msg's methodid will be initialized by NewObject.
If you want to use other msg mem, then you
can use the Set method on the method object.
*/

spmethod = NewObject(NULL, METHODCLASS, mobj_tags);


...
sp_msg.Pen = 0;
for (x = 0; x < 100; x ++)
{
    for (y = 0; y < 100; y ++)
    {
          sp_msg.X = x;
          sp_msg.Y = y;

          /* CallMethod() might be a macro */
          CallMethod(spmethod);

    }
};

-------------

For normal method invocation there are generally two approaches:

a) Hashing single methods.
b) Hashing interfaces, where an interface
is an array of methods (kinda like a AmigaOS library
jump table.)

a) is a bit faster than b)

In order to avoid having to hash the methods, all methods of a class
are in a single table. This includes the methods of all parent classes
so these can also be called directly (no need for DoSuperMethod()
in the dispatcher).

For this to work, we need:

- A method string -> ID to convert the method name to the method ID

- When you allocate a new class, you must allocate a table (size =
parentClass->numMethods + self->numMethods), copy the parent method
table and overwrite/add the own methods to the table

- The user uses a global variable in the classbase as the basic offset
to call methods of this class. The header file defines the offsets.

Interfaces work as usual but they store maybe only the real methodID
(one lookup more and one method object less).

Caveats:

- When a parent class changes its method table, then the child classes
won't notice. This could be avoided it we copy method objects.

- the DoMethod() stub must do an add for each method invokation but
no hashing.

Interfaces can be implemented like this:

- One hash table which contains tables of methods/method IDs and the
key is the interface ID.

- One hash table which contains method objects and the key is the
computed interface+method ID.

2) Noncentralized ID allocation.
-------------------------------
To allow for this, we avoid fixed integer
IDs (like in BOOPSI).
Instead we use global variables for
IDs.

Two ways to allocate IDs:

a)
Store interface IDs in global vars, and generate method IDs on the fly as
sum of interface ID and a offset ID for the method.

Example:

extern ULONG __OOPI_Root;

#define I_Root  __OOPI_Root
#define MIDX_Root_New           0
#define MIDX_Root_Dispose       1

#define M_Root_New      (I_Root + MIDX_Root_New)
#define M_Root_Dispose  (I_Root + MIDX_Root_Dispose)

Pro:
- Uses little mem.
- Relatively little CPU time used for initialization.

Con:
- uses more CPU time for initialization.

b)
Use one global var for each methodID.

Pro:
- Faster pr. call.
- Leaves more flexibility for future changes.
- Allows to specify each method as a string
in MakeClass().

Con:
- Uses more mem.
- Very hard to initialize


As for AttrIDs, there is a problem with using II), because then one can't
use switch() to parse AttrIDs.

Solution: Use methods for set & get of each attribute. which will use some
mem. Con: Clumsy.

To avoid the problems with init, we can use a global ClassBase (like a
library base) and use a public field in there to store the offsets. As for
attributes, a class should allocate a fixed number of IDs for its
attributes and store the base in the ClassBase. The header file then
contains offsets for each Attribute. In the class, you can very quickly
check if an attribute is part of the class:

    ULONG id = tag->ti_Tag - class->cl_attrBase;

    if (id < class->cl_numAttrs)
        ... handle own attributes
    else
        ... handle attributes of other classes

Location Transparent and Object Migration
-----------------------------------------

To allow these two, we use method objects. Method objects
contain the following attributes:

- A pointer to the method
- The method ID
- The class
- The object
- If several objects are created for the same method and the
method, for which they were created, changes, then all other
method objects change as well.
- The Method class contains a DoMethod() which is used to
call the method stored in the object.

By overloading DoMethod() of the Method class, you can implement
thread safe objects, migrate objects to other computers, etc.


Exceptions
----------

The easy part is to create the Exception class and the objects.

Now the hard part:

- How do I get the position information (stack frames, function names,
local variables (?), line numbers, filename) ?

- How do I "throw" an exception ?

- How do I catch an exception ?


On Unix, I can use GNU C extensions for most things. With a special
assembler, it should be possible to create position information and update
the call stack.

To throw and catch an exception, I can use a global array which contains
jmp_bufs and catched exceptions and throw just calls a global function
which looks through the array for a matching catch and does a longjmp()
to the place. We can use macros to throw and catch exceptions.


On Amiga, we might be able to use similar techniques. The globals
must be stored in the ETask info.

Thread safe method invocation.
-----------------------------

This can be done through the use of proxy objects
and IPC, where the server runs inside one task,
an the clients (other tasks) get proxies to the object
residing inside the server.
The server will the Wait() for incoming
methods and execute them on the proxied
object. Currently the system is using Exec messages,
and since AROS has no MP (yet), I only have to pass a
pointer.


A problem that has to be addressed here is
deadlocks. Say task A has object a, and task B
has object b.

Now, A calls a method on b, but the method calls another
method on a. Since A will be Wait()ing for a response from
B, then it can't execute the method, and both
tasks will be waiting forever.