"

 COPYRIGHT (c) 1991 by Claus Gittinger

	      All Rights Reserved

 This software is furnished under a license and may be used

 only in accordance with the terms of that license and with the

 inclusion of the above copyright notice.   This software may not

 be provided or otherwise made available to, or used by, any

 other person.  No title to or ownership of the software is

 hereby transferred.

"

ArrayedCollection variableSubclass:#WeakArray

	instanceVariableNames:'dependents'

	classVariableNames:'RegistrationFailedSignal AlreadyInitialized'

	poolDictionaries:''

	category:'Collections-Weak'

!

!WeakArray class methodsFor:'documentation'!

copyright

"

 COPYRIGHT (c) 1991 by Claus Gittinger

	      All Rights Reserved

 This software is furnished under a license and may be used

 only in accordance with the terms of that license and with the

 inclusion of the above copyright notice.   This software may not

 be provided or otherwise made available to, or used by, any

 other person.  No title to or ownership of the software is

 hereby transferred.

"

!

documentation

"

    WeakArrays can be used to trace disposal of objects; in contrast to other

    objects, references by WeakArrays will NOT keep an object from being

    garbage collected.

    Instead, whenever an object kept in a WeakArray dies, its entry is zeroed,

    and the WeakArray is informed by the storage manager. The WeakArray itself

    then informs possible dependents via the dependency mechanism.

    WeakArrays are used to track disposal of objects which keep external

    world resources. For example, FileStreams must close their underlying

    file when disposed (otherwise you could run out of OS filedescriptors).

    This can be done by keeping the FileStream objects in a weakArray, and

    keep a parallel array of filedescriptors. Whenever a fileStream is

    freed, search both arrays for an index where the stream is zero, but the

    filedescriptor is non-nil. Then close that file, and nil the filedescriptor

    entry. Notice, that there is a class (Registry) which does exactly this in

    a more programmer friendly way.

    Another application is caching of data: keep it in a weakArray, so the

    data in that cache will not be unreclaimable due to being cached.

    (for example, the ResourcePack class uses a WeakArray to cache recently

     used resource data for a while).

    The way in which weakArrays get informed by the runtime system is via

    an interrupt (DisposeInterrupt) which is first sent to the disposeHandler

    (typically ObjectMemory). ObjectMemory then takes the required steps to

    notify all weakArrays via the #lostPointer message.

    The reason for not sending messages directly from the VM is to make it 

    possible to run the finalization code at lower priority or from another class. 

    Also, as a side effect, it is possible to delay finalization by blocking 

    interrupts. (thus, the actual sending of the #lostPointer message is under

    control of Smalltalk code, which is modifyable).

    A weakArray notifies its dependents via normal dependency notfications.

    [hint:] 

	WeakArray handling adds small some overhead to the VM 

	(each weakarray is scanned after each GC). 

	It is uncertain, if the current mechanism works well

	with (say) ten-thousands of weakArrays.

	We had the system running with >2000 weakArrays, some being quite

	big for a while and had a few percent of added gc time.

	The system as delivered creates between 50 and 100 weakArrays,

	but with many dependents, this number may grow.

	If you need the dependency mechanism on a huge number of objects,

	consider adding a (non-weak) dependents field to your class

	- take the implementation of Model as a guide (or subclass them

	from Model).

    As a possible option, we could perform the weakArray scanning only in

    the oldSpace reclamation code - this would remove most of the overhead,

    but will lead to much longer delayed finalization .... we will see.

    [instance variables:]

	dependents                  get informed via #change notifiction 

				    that the weakArray has lost pointers.

				    Having the dependents here is an optimization.

    [class variables:]

	RegistrationFailedSignal    raised if a weakArray cannot be

				    registered by the VM. This only happens,

				    if the VM has to resize its internal tables

				    and is running out of malloc-memory.

    [memory requirements:]

	OBJ-HEADER + (size * ptr-size) + ptr-size

		   + sizeof(dependents-collection)

    [author:]

	Claus Gittinger

    [See also:]

	Array WeakIdentitySet WeakIdentityDictionary Registry

	Model

"

! !

!WeakArray class methodsFor:'initialization'!

initialize

    "setup the private signal"

    RegistrationFailedSignal isNil ifTrue:[

	RegistrationFailedSignal := ErrorSignal newSignalMayProceed:true.

	RegistrationFailedSignal nameClass:self message:#registrationFailedSignal.

	RegistrationFailedSignal notifierString:'weakArray registration failed'.

    ]

! !

!WeakArray class methodsFor:'instance creation'!

basicNew:size

    "return a new weakArray with size slots"

    "This is a kludge: I would like to set the WEAK-flag in the classes

     initialize method, but (currently) the order in which the class-initialize

     methods are called is not defined ...

     ... therefore it could happen, that a WeakArray is used by other

     classes initialize method BEFORE this method is evaluated. 

     To avoid this, the WEAK bit in the class is set here, when the very first 

     WeakArray is created.

     [Actually, the cleanest solution was to add another class definition

      keyword (such as #variableWeakSubclass:) which sets the weak flag right

      from the start.]"

    AlreadyInitialized isNil ifTrue:[

	self flags:(Behavior flagWeakPointers).

	AlreadyInitialized := true

    ].

    ^ (super basicNew:size) registerAsWeakArray

    "Modified: 25.1.1997 / 13:17:22 / cg"

! !

!WeakArray class methodsFor:'queries'!

isBuiltInClass

    "return true if this class is known by the run-time-system.

     Here, true is returned for myself, false for subclasses."

    ^ self == WeakArray

! !

!WeakArray methodsFor:'GC registration'!

registerAsWeakArray

    "register the receiver in the VM - 

     i.e. tell the VM to nil disposed entries in the receiver

     and notify the disposeInterruptHandler whenever that happened."

    |ok|

%{

    OBJ __addShadowObject();

    ok = __addShadowObject(self, 0);

    if (ok == false) {

	/* 

	 * the behavior of __addShadowObject() in case of overflowing

	 * VM-table space can be controlled by the second argument:

	 *   if its 0, the weakObject is not registered, and false

	 *   is returned.

	 *   if its 1, the tables are reallocated, registration proceeds,

	 *   and true is returned.

	 * This allows for the caller to have an influence on the VM's

	 * shadow table allocation.

	 *

	 * If addShadowObject() returned false, too many shadow objects are

	 * already there. Then collect garbage to get rid of

	 * obsolete ones, and try again.

	 * Since a full collect is expensive, we try

	 * a scavenge first, doing a full collect only if 

	 * that does not help.

	 *

	 * THIS MAY OR MAY NOT BE A GOOD IDEA: although it reduces

	 * the number of shadow objects that have to be

	 * processed at GC time, it may create a long delay here,

	 * at shadow object creation time.

	 * Dont know which is better ...

	 */

	__nonTenuringScavenge(__context);

	ok = __addShadowObject(self, 0);

	if (ok == false) {

	    /* 

	     * hard stuff - need full collect

	     * if this is the very first GC, assume that we are in

	     * the startup phase (where all weak stuff is allocated). 

	     * Then do no GC.

	     * Heuristics showed, that this GC does not find much ...

	     */

	    if ((__garbageCollectCount() != 0)

	     || (__incrementalGCCount() != 0)) {

		__markAndSweepIfUseful(__context);

		ok = __addShadowObject(self, 0);

	    }

	    if (ok == false) {

		/*

		 * mhmh - it seems that there are really many shadow 

		 * objects around - force creation

		 */

		ok = __addShadowObject(self, 1);

		if (ok == false) {

		    /* 

		     * no chance - something must be wrong 

		     * lets fall into the exception and see.

		     */

		}

	    }

	}

    }

%}.

    ok ifFalse:[

	"

	 the VM was not able to register the new weakArray

	 This can only happen, if the VM has to resize its tables,

	 and a malloc request failed. Usually, this smells like big

	 trouble being on the way (soon running out of memory in

	 other places as well).

	 Configure your OS for more swap space.

	"

	^ RegistrationFailedSignal raiseRequestWith:self

    ]

! !

!WeakArray methodsFor:'accessing'!

at:index

    "return the indexed instance variable with index, anInteger.

     Reimplemented here for IGC readBarrier. 

     (You dont have to understand this.)"

    ^ self basicAt:index

    "Modified: 25.1.1997 / 13:41:25 / cg"

!

at:index put:someObject

    "store someObject in the weakArray at some index."

    ^ self basicAt:index put:someObject

!

basicAt:index

    "return the indexed instance variable with index, anInteger.

     Reimplemented here for IGC readBarrier. 

     (You dont have to understand this.)"

%{  /* NOCONTEXT */

    REGISTER int indx;

    REGISTER unsigned int nIndex;

    OBJ el;

    if (__isSmallInteger(index)) {

	indx = __intVal(index) - 1;

	if (indx >= 0) {

	    nIndex = __BYTES2OBJS__(__qSize(self) - OHDR_SIZE);

	    indx += __intVal(__ClassInstPtr(__qClass(self))->c_ninstvars);

	    if (indx < nIndex) {

		el = __InstPtr(self)->i_instvars[indx];

		if (__isNonNilObject(el)) {

		    el = __WEAK_READ__(self, el);

		}

		RETURN (el);

	    }

	}

    }

%}.

    ^ super basicAt:index

!

basicAt:index put:someObject

    "store someObject in the weakArray at some index."

%{  /* NOCONTEXT */

    REGISTER int indx;

    REGISTER unsigned int nIndex;

    if (__isSmallInteger(index)) {

	indx = __intVal(index) - 1;

	if (indx >= 0) {

	    nIndex = __BYTES2OBJS__(__qSize(self) - OHDR_SIZE);

	    indx += __intVal(__ClassInstPtr(__qClass(self))->c_ninstvars);

	    if (indx < nIndex) {

		REGISTER OBJ el = someObject;

		__InstPtr(self)->i_instvars[indx] = el;

		if (__isNonNilObject(el)) {

		    __STORE(self, el);

		    __WEAK_WRITE__(self, el);

		}

		RETURN (el);

	    }

	}

    }

%}.

    "/ for the error-message ...

    ^ super basicAt:index put:someObject

! !

!WeakArray methodsFor:'copying'!

postCopy

    "copying alone does not really help - we have to tell

     the VM, that there is a new WeakArray around ...

     Q: who copies weakArrays ?"

    dependents := nil.

    self registerAsWeakArray.

!

skipInstvarIndexInDeepCopy:index

    "a helper for deepCopy; only indices for which this method returns

     false are copied in a deep copy."

    index == 1 ifTrue:[

        ^ true "/ skip dependents

    ].

    ^ false

! !

!WeakArray methodsFor:'dependents access'!

addDependent:anObject

    "make the argument, anObject be a dependent of the receiver"

    |wasBlocked|

    wasBlocked := OperatingSystem blockInterrupts.

    [

        |deps|

        deps := dependents.

        "/

        "/ store the very first dependent directly in

        "/ the dependents instVar

        "/

        (deps isNil and:[anObject isCollection not]) ifTrue:[

            dependents := anObject

        ] ifFalse:[

            "/

            "/ store more dependents in the dependents collection

            "/

            deps isCollection ifTrue:[

                deps add:anObject

            ] ifFalse:[

                deps == anObject ifFalse:[

                    deps isNil ifTrue:[

                        dependents := (IdentitySet with:anObject)

                    ] ifFalse:[

                        dependents := (IdentitySet with:deps with:anObject)

                    ]

                ]

            ]

        ]

    ] valueNowOrOnUnwindDo:[

        wasBlocked ifFalse:[

            OperatingSystem unblockInterrupts

        ]

    ]

    "Modified: 8.1.1997 / 23:40:30 / cg"

!

dependents 

    "return the dependents of the receiver"

    dependents isNil ifTrue:[^ #()].

    dependents isCollection ifTrue:[

	^ dependents

    ].

    ^ IdentitySet with:dependents

    "Modified: / 27.10.1997 / 19:37:16 / cg"

!

dependents:aCollection

    "set the dependents of the receiver"

    |dep|

    aCollection size == 1 ifTrue:[

	dep := aCollection first.

	dep isCollection ifFalse:[

	    dependents := aCollection first.

	    ^ self

	]

    ].

    dependents := aCollection

!

dependentsDo:aBlock

    "evaluate aBlock for all of my dependents"

    |deps|

    deps := dependents.

    deps notNil ifTrue:[

	deps isCollection ifTrue:[

	    deps do:aBlock

	] ifFalse:[

	    aBlock value:deps

	]

    ]

!

removeDependent:anObject

    "make the argument, anObject be independent of the receiver"

    |wasBlocked|

    "/ must do this save from interrupts, since the dependents collection

    "/ is possibly accessed from multiple threads.

    "/ Used to use #valueUninterruptably here; inlined that code for slightly

    "/ faster execution.

    wasBlocked := OperatingSystem blockInterrupts.

    [

	|deps sz dep|

	deps := dependents.

	deps notNil ifTrue:[

	    deps isCollection ifTrue:[

		deps remove:anObject ifAbsent:[].

		(sz := deps size) == 0 ifTrue:[

		    dependents := nil

		] ifFalse:[

		    sz == 1 ifTrue:[

			(dep := deps first) isCollection ifFalse:[

			    dependents := dep

			]

		    ]

		]

	    ] ifFalse:[

		deps == anObject ifTrue:[

		    dependents := nil

		]

	    ]

	]

    ] valueNowOrOnUnwindDo:[

	wasBlocked ifFalse:[

	    OperatingSystem unblockInterrupts

	]

    ]

! !

!WeakArray methodsFor:'enumerating'!

do:aBlock

    "evaluate the argument, aBlock for each element in the collection.

     - reimplemented for IGC readBarrier. 

     You dont have to understand this."

%{

    REGISTER int index;

    unsigned int nIndex;

    static struct inlineCache val = _ILC1;

    index = __intVal(__ClassInstPtr(__qClass(self))->c_ninstvars);

    nIndex = __BYTES2OBJS__(__qSize(self) - OHDR_SIZE);

    if (__isBlockLike(aBlock)

     && (__BlockInstPtr(aBlock)->b_nargs == __MKSMALLINT(1))) {