Smalltalk80LanguageImplementation:Chapter 30: Difference between revisions

From 흡혈양파의 번역工房
Jump to navigation Jump to search
(SMALLTALK80 Chapter 30 Formal Specification of the Object Memory 페이지 추가)
 
(형식 오류 수정)
Line 757: Line 757:
===A Space-efficient Reference-counting Collector===
===A Space-efficient Reference-counting Collector===


The traversal algorithm outlined above is recursive and, therefore, must use a stack in its execution. To guard against stack overflow, the depth of the stack must be greater than the longest chain of pointers in memory. This requirement is difficult to satisfy when memory space is limited. To guarantee that enough space is available, the pointer chain itself can be used as the stack. If object A references object B from A's ith field, and object B references object C from B's jth field, and object C references another object from C's kth field, and so on, the pointer chain can be represented as A.i ➛ B.j ➛ C.k ➛... (Figure 30.9a). To use the pointer chain as a stack for the recursion of the traversal algorithm, the chain is temporarily reversed to ...➛ C.k ➛ B.j ➛ A.i so that each field in the chain points to its predecessor instead of to its successor (Figure 30.9b).
The traversal algorithm outlined above is recursive and, therefore, must use a stack in its execution. To guard against stack overflow, the depth of the stack must be greater than the longest chain of pointers in memory. This requirement is difficult to satisfy when memory space is limited. To guarantee that enough space is available, the pointer chain itself can be used as the stack. If object A references object B from A's ith field, and object B references object C from B's j<sup>th</sup> field, and object C references another object from C's k<sup>th</sup> field, and so on, the pointer chain can be represented as A.i ➛ B.j ➛ C.k ➛... (Figure 30.9a). To use the pointer chain as a stack for the recursion of the traversal algorithm, the chain is temporarily reversed to ...➛ C.k ➛ B.j ➛ A.i so that each field in the chain points to its predecessor instead of to its successor (Figure 30.9b).




Line 810: Line 810:




Note that the last statement of the method restores the pointer chain to get B.j again pointing to C instead of to A. It is easy to do so when returning to B from processing C, because object pointer of C can simply be stored in the jth field of B. One might think that step unnecessary, because B is being deallocated. However, the same traversal algorithm can be used by a marking collector in which B is not being deallocated.
Note that the last statement of the method restores the pointer chain to get B.j again pointing to C instead of to A. It is easy to do so when returning to B from processing C, because object pointer of C can simply be stored in the j<sup>th</sup> field of B. One might think that step unnecessary, because B is being deallocated. However, the same traversal algorithm can be used by a marking collector in which B is not being deallocated.


<syntaxhighlight lang="smalltalk">
<syntaxhighlight lang="smalltalk">

Revision as of 11:32, 5 May 2014

Chapter 30 Formal Specification of the Object Memory

Formal Specification of the Object Memory

The two major components of any Smalltalk-80 implementation are the bytecode interpreter and the object memory. Chapters 28 and 29 described an implementation of the bytecode interpreter. This chapter describes an implementation of the object memory. The function of the object memory is to create, store, and destroy objects, and to provide access to their fields.


Memory-management systems fall into two major categories, realmemory implementations and virtual-memory implementations. In a real-memory implementation, all the objects in the environment reside in primary memory that is directly addressable by the program. In a virtual-memory implementation, objects reside in more than one level of a memory hierarchy and must be shuffled among the various levels during execution. This chapter describes the design of RealObjectMemory, an object memory for a real-memory Smalltalk-80.


Although Smalltalk can be implemented on computers of any word size, this presentation will be simplified by several assumptions in the standard algorithms. The routines of RealObjectMemory assume

  • that there are eight bits in a byte,
  • that there are two bytes in a word,
  • that the more significant byte of a word precedes the less significant byte, and
  • that the target computer is word addressed and word indexed.


Moreover, the routines assume that the address space is partitioned into 16 or fewer segments of 64K (65,536) words apiece. The standard algorithms can be systematically changed to adapt them to hardware with different properties. The routines of RealObjectMemory deal almost exclusively with 16-bit integers, as would a machine-language implementation.


To access locations in the address space of the host machine, machine language implementations use load and store instructions. In RealObjectMemory, the load and store instructions are symbolized by messages to an instance of RealWordMemory whose name is wordMemory. The protocol of RealWordMemory is shown below

RealWordMemory instance protocol
segment: s word: w Return word w of segment s
segment: s word : put: value Store value into word w of segment s; return value.
segment: s word: w byte: byteNumber Return byte byteNumber of word w of segment s.
segment: s word: w byte: byteNumber put: value Store value into byte byteNumber of word w of segment s; return value.
segment: s word: w bits: firstBitIndex to: lastBitIndex Return bits firstBitIndex to lastBitIndex of word w of segment s.
segment: s word: w bits: firstBitIndex to: lastBitIndex put: value Store value into bits firstBitIndex to lastBitIndex of word w of segment s; return value.


When it is necessary to distinguish the two bytes of a word, the left (more significant) byte will be referred to with the index 0 and the right (less significant) byte with the index 1. The most significant bit in a word will be referred to with the index 0 and the least significant with the index 15. Note that self is an instance of class RealObjectMemory in all routines of this chapter.


The most important thing about any implementation of the object memory is that it conform to the functional specification of the object memory interface given in Chapter 27. This chapter describes a range of possible implementations of that interface. In particular, simple versions of some routines are presented early in the chapter and refined versions are presented later as the need for those refinements becomes clear. These preliminary versions will be flagged by including the comment, "**Preliminary Version**", on the first line of the routine.


Heap Storage

In a real-memory implementation of Smalltalk, all objects are stored in an area called the heap. A new object is created by obtaining space to store its fields in a contiguous series of words in the heap. An object is destroyed by releasing the heap space it occupied. The format of an allocated object in the heap is shown in Figure 30.1. The actual data of the object are preceded by a two-word header. The size field of the header indicates the number of words of heap that the object occupies, including the header. It is an unsigned 16-bit number, and can range from 2 up to 65,536.

그림 30-1


When memory is segmented, it is Usually convenient for a Smalltalk-80 implementation to divide the heap into heap segments, each in a different memory segment. As stated earlier, the routines in this chapter assume that the target computer is segmented into address spaces of 65,536 words.

HeapSegmentCount The number of heap segments used in the implementation.
FirstHeapSegment The index of the first memory segment used to store the heap.
LastHeapSegment The index of the last memory segment used to store the heap (FirstHeapSegment + HeapSegmentCount - 1).
HeapSpaceStop The address of the last location used in each heap segment.
HeaderSize The number of words in an object header (2).
Heap Related Constants


그림 30-2


Compaction

Suppose for a moment that an object once allocated never changes its location in the heap. To allocate a new object, a space between existing objects must be found that is large enough to hold the new object. After a While, the memory "fragments" or "checkerboards." That is, an allocation request is bound to arrive for an amount of space smaller than the total available memory but larger than any of the disjoint pieces (Figure 30.2a). This can occur even if there is a large amount of available space and a relatively small allocation request.


Fragmentation cannot be tolerated in an interactive system that is expected to preserve a dynamic environment for hundreds of hours or more without reinitialization. Therefore when memory fragments, it must be compacted. Memory is compacted by moving all objects that are still in use towards one end of the heap, squeezing out all the free space between them and leaving one large unallocated block at the other end (see Figure 30.2b).


Each heap segment is compacted separately. Even on a linearly-addressed machine it is preferable to segment a large heap to reduce the duration of each compaction.


The Object Table

When an object is moved during compaction, all pointers to its heap memory must be updated. If many other objects contain pointers directly to the old location, then it is time-consuming on a sequential computer to find and update those references to point to the new location. Therefore to make the pointer update inexpensive, only one pointer to an object's heap memory is allowed. That pointer is stored in a table called the object table. All references to an object must be indirected through the object table. Thus, the object pointers found in Smalltalk objects are really indices into the object table, in which pointers into the heap are in turn found (see Figure 30.3).


Indirection through the object table provides another benefit. The number of objects of average size Z addressable by an n-bit pointer is on the order of 2n instead of 2n/Z. In our experience, objects average 10 words in size (Z~10), so a significant gain in address space can be realized by indirection.


Throughout the object table, abandoned entries can occur that are not associated with any space on the heap. These entries are called free entries and their object pointers are called free pointers. It is easy to recycle a free entry, because all object table entries are the same size. Compaction of the object table is difficult and generally unnecessary, so it is not supported.

그림 30-3


Although the heap is segmented, the object table is stored in a single segment so that an object pointer can be 16 bits and thus fit in one word. Consequently, the number of objects that can be addressed in real memory is limited to the number of object table entries that can fit in one segment. A common arrangement is for each object table entry to occupy two words and for the entire table to occupy 64K words or less, yielding a maximum capacity of 32K objects.


Object Pointers

An object pointer occupies 16 bits, apportioned as in Figure 30.4.

그림 30-4


When the low-order bit of the object pointer is 0, the first 15 bits are an index into the object table. Up to 215(32K) objects can be addressed. When the low-order bit of the object pointer is 1, the first 15 bits are an immediate signed integer, and no additional space in the object table or the heap is utilized. The benefit of giving special treatment to integers in the range ±214 is that they come and go with high frequency during arithmetic and many other operations. The cost of their efficient representation is the number of tests the interpreter must perform to distinguish object pointers of small integers from object pointers of other objects.


The isIntegerObject: routine tests the low order bit of objectPointer to determine whether the rest of the pointer is an immediate integer value rather than an object table index.

isIntegerObject: objectPointer
    (objectPointer bitAnd: 1) = 1


Every other object-access routine requires that its object pointer argument really be an object table index. The cantBeIntegerObject: routine is used to trap erroneous calls. If the hardware, the bytecode interpreter, and the object memory manager are bug free, then this error condition is never encountered.

cantBeIntegerObject: objectPointer
    (self isIntegerObject: objectPointer)
        ifTrue: [Sensor notify: 'A small integer has no object table entry']


Object Table Entries

The format of an object table entry is shown in Figure 30.5. If the free entry bit is on, then the entry is free. If the free entry bit is off, then the four segment bits select a heap segment and the 16 location bits locate the beginning of the space in that segment that is owned by the object table entry. The count field, the odd length bit (O), and the pointer fields bit will be explained later in the chapter.

그림 30-5


ObjectTableSegment The number of the memory segment containing the object table.
ObjectTableStart The location in objectTableSegment of the base of the object table.
ObjectTableSize The number of words in the object table (an even number 수식 64K).
HugeSize The smallest number that is too large to represent in an eight-bit count field; that is, 256.
NilPointer The object table index of the object nil
Object Table Related Constants


The following set of routines accesses the first word of object table entries in four different ways: loading the whole word, storing the whole word, loading a bit field, and storing a bit field. These routines in turn utilize routines of wordMemory, an instance of RealWordMemory. They assume that objectPointer is expressed as an even word offset relative to objectTableStart, the base of the object table in segment objectTableSegment. Note that ot is an abbreviation for "object table."

ot: objectPointer
    self cantBeIntegerObject: objectPointer.
    wordMemory segment: ObjectTableSegment
        word: ObjectTableStart + objectPointer

ot: objectPointer put: value
    self cantBeIntegerObject: objectPointer
    wordMemory segment: ObjectTableSegment
        word: ObjectTableStart + objectPointer
        bits: firstBitIndex
        to: lastBitIndex

ot: objectPointer bits: firstBitIndex to: lastBitIndex put: value
    self cantBeIntegerObject: objectPointer. 
    wordMemory segment: ObjectTableSegment
        word: ObjectTableStart + objectPointer 
        bits: firstBitIndex
        to: lastBitIndex
        put: value


The following 12 object-access subroutines load or store the various fields of the object table entry of objectPointer.

countBitsOf: objectPointer
    self ot: objectPointer bits: 0 to: 7
countBitsOf: objectPointer put: value
    self ot: objectPointer bits: 0 to: 7 put: value
oddBitOf: objectPointer
    self ot: objectPointer bits: 8 to: 8
oddBitOf: objectPointer put: value
    self ot: objectPointer bits: 8 to: 8 put: value
pointerBitOf: objectPointer
    self ot: objectPointer bits: 9 to: 9
pointerBitOf: objectPointer put: value
    self ot: objectPointer bits: 9 to: 9 put: value
freeBitOf: objectPointer
    self ot: objectPointer bits: 10 to: 10
freeBitOf: objectPointer put: value
    self ot: objectPointer bits: 10 to: 10 put: value
segmentBitsOf: objectPointer
    self ot: objectPointer bits: 12 to: 15
segmentBitsOf: objectPointer put: value
    self ot: objectPointer bits: 12 to: 15 put: value

locationBitsOf: objectPointer
    self cantBeIntegerObject: objectPointer. 
    twordMemory segment: ObjectTableSegment
        word: ObjectTableStart + objectPointer + 1

locationBitsOf: objectPointer put: value
    self cantBeIntegerObject: objectPointer.
    wordMemory segment: ObjectTableSegment
        word: ObjectTableStart + objectPointer + 1
        put: value


For objects that occupy a chunk of heap storage (those whose free bit is 0), the following four object-access subroutines load or store words or bytes from the chunk.

heapChunkOf: objectPointer word: offet
    wordMemory segment: (self segmentBitsOf: objectPointer)
        word: ((self locationBitsOf: objectPointer) + offset)
heapChunkOf: objectPointer word: offset put: value
    wordMemory segment: (self segmentBitsOf: objectPointer)
        word: ((self locationBitsOf: objectPointer) + offset)
        put: value
heapChunkOf: objectPointer byte: offset
    wordMemory segment: (self segmentBitsOf: objectPointer)
        word: ((self locationBitsOf: objectPointer) + (offset // 2))
        byte: (offset \\ 2)
heapChunkOf: objectPointer byte: offset put: value
    wordMemory segment: (self segmentBitsOf: objectPointer)
        word: ((self locationBitsOf: objectPointer) + (offset // 2))
        byte: (offset \\ 2) put: value


The next four object-access subroutines are more specialized in that they load or store words of the object header.

sizeBitsOf: objectPointer
    self heapChunkOf: objectPointer word: 0
sizeBitsOf: objectPointer put: value
    self heapChunkOf: objectPointer word: 0 put: value
classBitsOf: objectPointer
    self heapChunkOf: objectPointer word: 1
classBitsOf: objectPointer put: value
    self heapChunkOf: objectPointer word: 1 put: value


The remaining two object-access subroutines are functionally identical to sizeBitsOf: in the versions shown below. Later in this chapter, refinements to the object-memory manager will require new versions of both of these subroutines that will return something different from the object size in certain cases. For that reason, these methods are marked "preliminary."

lastPointerOf: objectPointer "**Preliminary Version**"
    self sizeBitsOf: objectPointer
spaceOccupiedBy: objectPointer "**Preliminary Version**"
    self sizeBitsOf: objectPointer


Unallocated Space

All free entries in the object table are kept on a linked list headed at the location named freePointerList. The link from one free entry to the next is an object pointer in its location field (see Figure 30.6).

그림 30-6


Unallocated space in the heap is grouped into free chunks (contiguous blocks) of assorted sizes and each of those free chunks is assigned an object table entry. Free chunks are linked together on lists, each containing chunks of the same size. The link from one free chunk to the next is in its class field (Figure 30.7). To keep the table of list heads small, all free chunks bigger than 20 words are linked onto a single list.

FreePointerList The location of the head of the linked list of free object table entries.
BigSize The smallest size of chunk that is not stored on a list whose chunks are the same size. (The index of the last free chunk list).
FirstFreeChunkList The location of the head of the linked list of free chunks of size zero. Lists for chunks of larger sizes are stored in contiguous locations following FirstFreeChunkkist. Note that the lists at FirstFreeChunkList and FirstFreeChunkList + 1 will always be empty since all chunks are at least two words long.
LastFreeChunkList The location of the head of the linked list of free chunks of size BigSize or larger.
NonPointer Any sixteen-bit value that cannot be an object table index, e.g., 216-1
Free Space Related Constants


A separate set of free chunk lists is maintained for each heap segment, but only one free pointer list is maintained for the object table. Note that the object table entry associated with a "freechunk" is not a "free entry." It is not on the free pointer list, and its free entry bit is not set. The way a free chunk is distinguished from an allocated chunk is by setting the count field of the object table entry to zero for a free chunk and to nonzero for an allocated chunk.

그림 30-7


The following four routines manage the free pointer list headed at freePointerList in segment objectTableSegment. The first two routines simply load and store the list head.

headOfFreePointerList
    wordMemory segment: ObjectTableSegment
        word: FreePointerList

headOfFreePointerListPut: objectPointer
    wordMemory segment: ObjectTableSegment
        word: FreePointerList
        put: objectPointer


The routine toFreePointerListAdd: adds a free entry to the head of the list.

toFreePointerListAdd: objectPointer
    self locationBitsOf: objectPointer
        put: (self headOfFreePointerList)
    self headOfFreePointerListPut: objectPointer


The routine removeFromFreePointerList removes the first entry from the list and returns it; if the list was empty, it returns nil. The distinguished value NonPointer signifies the end of a linked list. A good value for NonPointer is 수식, a value that is easily detected on most computers and that cannot be confused with an actual object table entry address because it is an odd number.

removeFromFreePointerList
    | objectPointer |
    objectPointer  self headOfFreePointerList.
    objectPointer = NonPointer ifTrue: [nil].
    self headOfFreePointerListPut: (self locationBitsOf: objectPointer).
    objectPointer


The following routines manage the free-chunk lists headed at FirstFreeChunkList + 2 through LastFreeChunkList of each heap segment. Their behavior is exactly analogous to that of the routines above. The first three routines work in the segment specified or implied by their second parameter. The fourth routine works in the segment specified by the register currentSegment.

headOfFreeChunkList: size inSegment: segment
    wordMemory segment: segment
        word: FirstFreeChunkList + size

headOfFreeChunkList: size inSegment: segment put: objectPointer
    wordMemory segment: segment
        word: FirstFreeChunkList + size
        put: objectPointer

toFreeChunkList: size add: objectPointer
    | segment |
    segment  self segmentBitsOf: objectPointer.
    self classBitsOf: objectPointer
        put: (self headOfFreeChunkList: size inSegment: segment).
    self headOfFreeChunkList: size
        inSegment: segment
        put: objectPointer

removeFromFreeChunkList: size
    | objectPointer secondChunk |
    objectPointer  self headOfFreeChunkList: size
        inSegment: currentSegment.
    objectPointer = NonPointer ifTrue: [nil].
    secondChunk  self classBitsOf: objectPointer.
    self headOfFreeChunkList: size
        inSegment: currentSegment
        put: secondChunk.
    objectPointer


The routine resetFreeChunkList:inSegment: resets the specified freechunk list to an empty list.

resetFreeChunkList: size inSegment: segment
    self headOfFreeChunkList: size
        inSegment: segment
        put: NonPointer


Allocation and Deallocation

To allocate an object, an entry is obtained from the object table and sufficient space for the header and data is obtained from some heap segment. The heap segment in which space is found is called the current segment. It becomes the first segment in which to look for space to allocate the next object. The only register required by the object memory holds the index of the current segment.

currentSegment The index of the heap segment currently being used for allocation.
Registers of the Object Memory


To allocate a "large" object requiring n words of heap space (n > = BigSize), the list beginning at LastFreeChunkList in the current segment is searched for a free chunk whose size is either n words or at least n + headerSize words. If the free chunk found is larger than n words, it is subdivided and only n of the words are used to satisfy the allocation request.


To allocate a "small" object requiring n words of heap space (headerSize < = n < BigSize), the list beginning at freeChunkLists + n is searched. If the list is nonempty, its first free chunk is removed and used for the new object. If the list is empty, the above algorithm for "large" objects is used.


If no chunk of sufficient size is found in the current segment, then the next segment is made current and the search continues there. The new current segment is compacted first to improve the chances of finding sufficient space. In a compacted segment, all the allocated objects are at one end and the (presumably large) space at the other end is all in one large chunk, the sole member of the list LastFreeChunkLists. If enough space is not found in any segment, execution is halted.


When an object is deallocated, its space is recycled on the list of free chunks of the appropriate size. However, to simplify the presentation in this chapter, the standard algorithms leave the unused part of any subdivided chunk on the list of big free chunks even if that part is small in size.


An Allocation Algorithm

The allocate:class: routine is presented below as an example of a simple allocation routine. It takes as parameters the size of the desired chunk (in words, including header) and the class of the object that chunk will represent. The actual allocation routine takes several other parameters and so the allocate:class: routine will be flagged as preliminary. A more complete routine, allocate:extra:class:, is presented in a later section and the actual routine used in the implementation, allocate:odd:pointer:extra:class:, is presented after that.

allocate: size class: classPointer "**Preliminary Version**"
    | object Pointer |
    objectPointer  self allocateChunk: size. "actually allocate"
    self classBitsOf: objectPointer put: classPointer. "fill in class"
    "initialize all fields to the object table index of the object nil"
    (headerSize to: size - 1) do:
        [ :i | self heapChunkOf: objectPointer word: i put: NilPointer].
    self sizeBitsOf: objectPointer put: size.
    "return the new object's pointer"
    objectPointer


The routine allocateChunk: either succeeds in its allocation task, or it reports an unrecoverable error. It uses a subroutine, attemptToAllocateChunk:, that either completes the job or returns nil if no space can be found.

allocateChunk: size "**Preliminary Version**"
    | objectPointer |
    objectPointer  self attemptToAllocateChunk: size.
    objectPointer isNil ifFalse: [objectPointer].
    self error: 'Out of memory'


The attemptToAllocateChunk: routine first tries to allocate in currentSegment, the segment currently targeted for allocations. It does so using the subroutine attemptToAllocateChunkInCurrentSegment:. If the subroutine fails (returns nil), then the routine compacts the next segment and retries the allocation there. This procedure continues until the original segment has been compacted and searched. If no space can be found anywhere, the routine returns nil. Note that it uses implementation-dependent constants: HeapsegmentCount, FirstHeapsegment, and LastHeapsegment.

attemptToAllocateChunk: size
    | objectPointer |
    objectPointer  self attemptToAllocateChunkInCurrentSegment: size. 
    objectPointer isNil ifFalse: [objectPointer].
    1 to: HeapSegmentCount do:
        [ :i |
            currentSegment  currentSegment + 1.
            currentSegment > LastHeapSegment
                ifTrue: [currentSegment  FirstHeapSegment]. 
            self compactCurrentSegment.
            objectPointer
                     self attemptToAllocateChunkInCurrentSegment: size. 
            objectPointer isNil ifFalse: [objectPointer]].
    nil


The attemptToAllocateChunkInCurrentSegment: routine searches the current heap segment's free-chunk lists for the first chunk that is the right size or that can be subdivided to yield a chunk of the right size. Because most objects are smaller than BigSize and most allocation requests can be satisfied by recycling deallocated objects of the desired size, most allocations execute only the first four lines of the routine.

attemptToAllocateChunkInCurrentSegment: size
    | objectPointer predecessor next availableSize excessSize newPointer | 
    size < BigSize
        ifTrue: [objectPointer  self removeFromFreeChunkList: size]. 
    objectPointer notNil
        ifTrue: [objectPointer]. "small chunk of exact size handy so use it" 
    predecessor  NonPointer.
                "remember predecessor of chunk under consideration" 
    objectPointer  self headOfFreeChunkList: LastFreeChunkList
                            inSegment: currentSegment.
    "the search loop stops when the end of the linked list is encountered"
    [objectPointer = NonPointer] whileFalse: 
        [availableSize  self sizeBitsOf: objectPointer.
            availableSize = size
                ifTrue: "exact fit --remove from free chunk list and return"
                    [next  self classBitsOf: objectPointer. 
                                    "the link to the next chunk"
                        predecessor = Non Pointer
                            ifTrue: "it was the head of the list; make the next item the head "
                                [self headOfFreeChunkList: LastFreeChunkList 
                                    inSegment: currentSegment put: next]
                            ifFalse: "it was between two chunks; link them together" 
                                [self classBitsOf: predecessor
                                    put: next]. 
                        objectPointer].
    "this chunk was either too big or too small; inspect the amount of variance "
    excessSize  availableSize-size. 
    excessSize > = HeaderSize
        ifTrue: "can be broken into two usable parts: return the second part" 
            ["obtain an object table entry for the second part"
                newPointer  self obtainPointer: size
                                    location: (self locationBitsOf: objectPointer)
                                                        + excessSize.
                newPointer isNil ifTrue: [nii].
                "correct the size of the first part (which remains on the free list)"
                self sizeBitsOf: objectPointer put: excessSize.
                newPointer]
        ifFalse: "not big enough to use; try the next chunk on the list"
            [predecessor  objectPointer.
                objectPointer  self classBitsOf: objectPointer]].
nil "the end of the linked list was reached and no fit was found"


The subroutine obtainPointer:location: used by the above routine obtains a free object table entry, zeroes its free entry bit as well as the rest of the first word of the entry, points the entry at the specified location, and sets the size field of the header to the specified size.

obtainPointer: size location: location
    | objectPointer |
    objectPointer  self removeFromFreePointerList.
    objectPointer isNil ifTrue: [nil].
    self ot: objectPointer put: 0.
    self segmentBitsOf: objectPointer put: currentSegment.
    self locationBitsOf: objectPointer put: location.
    self sizeBitsOf: objectPointer put: size.
    objectPointer


A Deallocation Algorithm

It is much simpler to deallocate an object than to allocate one. The chunk is recycled on a free-chunk list. The following routine expects the count field to have been reset to zero by a higher-level routine

deallocate: objectPointer "**Preliminary Version**"
    | space |
    space  self spaceOccupiedBy: objectPointer.
    self toFreeChunkList: (space min: BigSize)
        add: objectPointer


Note that this routine computes the space occupied by the object using spaceOccupiedBy: instead of sizeBitsOf:. The reason will become clear later in the chapter when spaceOccupiedBy: is redefined.


A Compaction Algorithm

The compactCurrentSegment routine invoked above by attemptToAllocateChunk: sweeps through a heap segment, massing all allocated objects together and updating their object table entries. For the benefit of subsequent allocation, it also links the object table entries reclaimed from the free chunk lists onto the free pointer list and creates a single free chunk from the unused portion of the heap segment. The algorithm for compactCurrentSegment will be presented shortly, after some preparatory discussion.


After a heap segment is compacted a number of times, relatively permanent objects sift to the bottom of the segment and most allocation and deallocation activity occurs nearer to the top. The segment consists of a densely packed region of allocated chunks, followed by a region of both allocated and free chunks. During compaction, chunks in the densely packed region never move, because there is no space beneath them to eliminate. Therefore, the compacter expends effort only on chunks above the first free chunk, whose location is referred to as lowWaterMark.


The abandonFreeChunksInSegment: routine computes lowWaterMark. It also finds all deallocated chunks, recycles their object table entries onto the list of free pointers using the subroutine releasePointer:, and changes their class fields to the distinguished value nonPointer. During the subsequent sweep, when the compacter encounters objects so marked it can recognize them as deallocated chunks.

abandonFreeChunksInSegment: segment
    | lowWaterMark objectPointer nextPointer |
    lowWaterMark  HeapSpaceStop. "first assume that no chunk is free"
    HeaderSize to: BigSize do: "for each free-chunk list"
        [ :size |
            objectPointer  self headOfFreeChunkList: size
                                    inSegment: segment.
            [objectPointer = NonPointer] whileFalse:
                [lowWaterMark  lowWaterMark
                                        min: (self locationBitsOf: objectPointer).
                    nextPointer  self classBitsOf: objectPointer.
                                            "link to next free chunk"
                    self classBitsOf: objectPointer put: NonPointer
                                            "distinguish for sweep"
                    self releasePointer: objectPointer.
                                            "add entry no free-pointer list"
                    objectPointer  nextPointer].
                self resetFreeChunkList: size inSegment: segment].
    lowWaterMark

releasePointer: objectPointer
    self freeBitOf: objectPointer put: 1.
    self toFreePointerListAdd: objectPointer


A heap segment is compacted by sweeping through it from bottom to top. Each allocated object is moved as far down in the segment as possible without overwriting other allocated objects. For each object moved, the corresponding object table entry is found and its location field is updated to point to the new location of the object.


It is by no means trivial to find the object table entry of an object encountered during a sweep of the heap segment. The representation of the object in the heap does not include a pointer back to the object table entry. To avoid the cost of such a backpointer for every object or making the compacter search the object table after every object is moved, a trick called "reversing pointers" is employed. During compaction, instead of the usual arrangement in which the object table entry points to the header in the heap, the header points temporarily to the object table entry.


Pointers are reversed before starting to sweep through a heap segment. The object table is scanned to find every in-use entry whose segment field refers to the segment being compacted and whose location field is above lowWaterMark. Each such entry points to the header of an object that is to be moved (Figure 30.8a). The pointer is then reversed, i.e., the object's own object pointer is stored in the first word of its header. This causes the header to point to the object table entry. By doing this, the size field of the header is overwritten.: To prevent losing the size, it is saved in the second word of the object table entry (Figure 30.8b). By doing that, the location field is overwritten, but that is of no consequence, because the compacter recomputes the object's heap location after the move.

그림 30-8


reverseHeapPointersAbove: lowWaterMark
    |size|
    0 to: ObjectTableSize-2 by: 2 do:
        [ :objectPointer |
            (self freeBitOf: objectPointer) = 0
                ifTrue: "the Object Table entry is in use"
                    [(self segmentBitsOf: objectPointer) = currentSegment
                        ifTrue: "the object is in this segment"
                            [(self locationBitsOf: objectPointer) < lowWaterMark
                                ifFalse: "the object will be swept"
                                    [size  self sizeBitsOf: objectPointer.
                                                        "rescue the size"
                                        self sizeBitsOf: objectPointer
                                            put: objectPointer. "reverse the pointer"
                                        self locationBitsOf: objectPointer
                                            put: size "save the size"]]]]


After all preparations for compaction are complete, the current heap segment is swept using the sweepCurrentSegmentFrom routine. It maintains two pointers into the segment, si (source index) and di (destination index). The pointer si points to the header of an object currently being considered for retention or elimination. The pointer di points to the location where that object will be moved if retained.

sweepCurrentSegmentFrom: lowWaterMark
    | si di objectPointer size space |
    si  di  lowWaterMark.
    [si < HeapSpaceStop]
        whileTrue: "for each object, si"
            [(wordMemory segment: currentSegment word: si + 1) = NonPointer
                ifTrue: "unallocated, so skip it"
                    [size  wordMemory segment: currentSegment word: si.
                        si  si + size]
                ifFalse: "allocated, so keep it, but move it to compact storage"
                    [objectPointer
                             wordMemory
                                    segment: currentSegment word: si.
                        size  self locationBitsOf: objectPointer.
                                    "the reversed sie"
                        self locationBitsOf: objectPointer
                            put: di. "point object table at new location"
                        self sizeBitsOf: objectPointer
                            put: size. "restore the size to its proper place"
                        si  si + 1. "skip the size"
                        di  di + 1. "skip the size"
                        2 to: (self spaceOccupiedBy: objectPointer) do:
                                "move the rest of the object"
                            [ :i |
                                wordMemory segment: currentSegment
                                            word: di
                                            put: (wordMemory segment:
                                                                currentSegment
                                                            word: si).
                                si  si + 1.
                                di  di + 1]]].
    di


Note that while pointers are reversed, it is impossible to access the heap memory of an object from its object table entry. Therefore the Smalltalk interpreter cannot run during compaction.


The compactCurrentSegment routine invokes the above routines in the proper order and then creates the single free chunk at the top of the heap segment.

compactCurrentSegment
    | lowWaterMark bigSpace |
    lowWaterMark  self abandonFreeChunksInSegment: currentSegment.
    lowWaterMark < HeapSpaceStop
        ifTrue:
            [self reverseHeapPointersAbove: lowWaterMark.
                bigSpace  self sweepCurrentSegmentFrom: lowWaterMark.
                self deallocate: (self obtainPointer:
                                            (HeapSpaceStop + 1 - bigSpace)
                                        location: bigSpace)]


If there are no free chunks within the segment when this routine is invoked, then it does not move any objects.


Garbage Collection

In Smalltalk, a new object is allocated explicitly (e.g., when the message new is sent to a class) but there is no explicit language construct that causes an object to be deallocated. Such a construct would be unsafe, because it could be used to deallocate an object even though "Mangling" references to it still existed in other objects. An environment containing dangling references would be inconsistent and would be likely to exhibit unintended behavior and to suffer unrecoverable errors.


Most noninteractive programming systems require explicit deallocation. The burden of avoiding dangling references is placed on the programmer. If a dangling reference arises, the programmer is expected to find the bug that created it, fix that bug, and restart the program. In an interactive environment like Smalltalk (as well as most LISP and APL systems), to require a restart because of a common bug would be unacceptable, since it could require the user to redo a potentially large amount of work.


Because there is no explicit deallocation in Smalltalk, the memory manager must identify objects that have become inaccessible and deallocate them automatically. This task is traditionally known as garbage collection. As compared with explicit deallocation, garbage collection entails a large performance penalty. The penalty is incurred because the computer must manage deallocation at execution time instead of relying on the programmer to have done so during coding time. However, the cost is well worth the reliability it adds to an interactive system.


There are two traditional approaches to identifying inaccessible objects in an object memory: marking and reference counting. A marking garbage collector performs an exhaustive search of memory for accessible objects and marks them all. Then it scans memory in search of objects that are unmarked and thus inaccessible and deallocates them. A reference-counting garbage collector maintains a count of how many references there are to each object from other objects. When the count of references to some object reaches zero, that object is known to be inaccessible, and the space it occupies can be reclaimed.


Reference-counting systems do not deal properly with so-called "cyclic structures." Such a structure is said to occur when an object references itself directly or when an object references itself indirectly via other objects that reference it. In a reference-counting system, when a cyclic structure becomes inaccessible to the program, it will have nonzero reference counts due to the intrastructure references. Therefore the memory manager doesn't recognize that the structure should be deallocated, and the objects that constitute the structure are not deallocated. These inaccessible objects waste space; but, unlike dangling references, they do not cause inconsistency in the environment.


Both reference counting and marking involve performance penalties on conventional computers. In a reference-counting system, the frequently performed operation of storing a reference to an object involves overhead for reference-count maintenance, so programs run significantly more slowly. In a marking garbage collector, an extensive search of memory must be performed whenever space is entirely depleted. Therefore, program execution is subject to relatively lengthy interruptions that can be quite annoying in an interactive system. Both approaches incur space overhead. In a reference-counting system, space must be provided to store reference counts. In a marking system, extra space must be allotted in the heap to allow garbage to accumulate between collections. Otherwise, collections occur too frequently.


The approach to garbage collection that should be used in a particular implementation of Smalltalk depends in part on the capacity of the hardware. If a relatively small amount of memory (e.g., 100 kilobytes) is available, a reference counting system is intolerable, because it can waste precious space by leaving inaccessible cyclic structures allocated. On the other hand, a marking collector is quite acceptable in these circumstances, in spite of the interruption that occurs when it is invoked, because when memory is small, the duration of the interruption can be so brief as to be imperceptible. If an abundant supply of memory (e.g., two megabytes) is available, the time it takes to mark all accessible objects can be so long as to be intolerable. On the other hand, there is enough space available that a moderate number of inaccessible objects can be tolerated.


The contrast between the two approaches is accentuated in a large virtual-memory system. Marking is even more costly because so much time is spent in random accesses to secondary memory. Reference counting is even less costly because unreclaimed cyclic structures simply migrate to secondary memory where wasted space is less bothersome. When memory is abundant, a reference-counting garbage collector is appropriate. However, Smalltalk programmers should take precautions to avoid the accumulation of an excessive number of inaccessible cyclic structures, otherwise even a large memory will be depleted. To break a cyclic structure before it becomes inaccessible, the program can change any pointer that participates in the cycle to nil.


The two approaches to garbage collection can be combined. References can be counted during normal operation and marking collections performed periodically to reclaim inaccessible cyclic structures. A combined approach is suitable for all but the smallest real-memory implementations. If a small-to-medium-size memory is available, a marking collection can be performed whenever compaction fails to recover enough space. If an abundant memory is available, marking collections can be performed nightly or at other convenient intervals.


A Simple Reference-counting Collector

In the reference-counting collector described in this chapter, the reference count of an object is recorded in the count field of its object table entry. If an object pointer is an immediate integer, it is its own only reference, so its reference count is not recorded explicitly. Reference counts are updated during store operations. When an object pointer referencing object P is stored into a location that formerly contained an object pointer referencing object Q, the count field of P is incremented and the count field of Q is decremented. Because the count field of an object table entry has only eight bits, it is possible for an incremented count to overflow. To facilitate overflow detection on most computers, the high order bit of the count field serves as an overflow bit. Once the count field reaches 128, it remains at that value and it will not increase or decrease. The algorithm for incrementing a reference count is

countUp: objectPointer
    | count |
    (self isIntegerObject: objectPointer)
        ifFalse:
            [count  (self countBitsOf: objectPointer) + 1.
                count < 129 ifTrue: [self countBitsOf: objectPointer put: count]].
    objectPointer


If the decremented reference count of an object reaches zero, then that object is deallocated. Before doing so, the count field of every object referenced from that object is decremented, because once the object is deallocated it will no longer reference those other objects. Note that this procedure recurs if any of the latter counts reach zero. A recursive procedure that can traverse the original object plus all the objects it references is expressed below as the routine forAllObjectsAccessibleFrom:suchThat:do:. This routine takes two procedural arguments represented by blocks, a predicate that decrements a count and tests for zero and an action that deallocates an object. Between evaluating the predicate and the action, the procedure's subroutine, forAllOtherObjectsAccessibleFrom:suchThat:do:, recursively processes every pointer in the object. The procedure is allowed to alter the count as a side effect, so the action argument must restore the count to zero in preparation for deallocation.

countDown: rootObjectPointer
    | count |
    (self isIntegerObject: rootObjectPointer)
        ifTrue: [rootObjectPointer]
        ifFalse: "this is a pointer, so decrement its reference count"
            [self forAllObjectsAccessibleFrom: rootObjectPointer
                suchThat:
                    "the predicate decrements the count and tests for zero"
                    [ :objectPointer |
                        count  (self countBitsOf: objectPointer) - 1.
                        count < 127
                            ifTrue: [self countBitsOf: objectPointer
                                put: count].
                        count = 0]
                do: "the action zeroes the count and deallocates the object"
                    [ :objectPointer |
                        self countBitsOf: objectPointer put: 0.
                        self deallocate: objectPointer]]


The traversal routine shown below first tests the predicate on the supplied object. It then invokes a subroutine that (1) recursively processes all objects referenced from within the supplied object that satisfy predicate, and (2) performs action on the supplied object.

forAllObjectsAccessibleFrom: objectPointer suchThat: predicate do: action
    (predicate value: objectPointer)
        ifTrue:
            [self forAllOtherObjectsAccessibleFrom: objectPointer
                suchTaht: predicate
                do: action]

forAllOtherObjectsAccessibleFrom: objectPointer suchThat: predicate do:action
    | next|
    1 to: (self lastPointerOf: objectPointer) - 1 do:
        [ :offset |
            next  self heapChunkOf: objectPointer word: offset.
            ((self isIntegerObject: next) = = false and: [predicate value: next])
                ifTrue: "it's a non-immediate object and it should be processed"
                    [self forAllOtherObjectsAccessibleFrom: next
                        suchThat: predicate
                        do: action]].
    "all pointers have been followed; now perform the action"
    action value: objectPointer.
    objectPointer


A Space-efficient Reference-counting Collector

The traversal algorithm outlined above is recursive and, therefore, must use a stack in its execution. To guard against stack overflow, the depth of the stack must be greater than the longest chain of pointers in memory. This requirement is difficult to satisfy when memory space is limited. To guarantee that enough space is available, the pointer chain itself can be used as the stack. If object A references object B from A's ith field, and object B references object C from B's jth field, and object C references another object from C's kth field, and so on, the pointer chain can be represented as A.i ➛ B.j ➛ C.k ➛... (Figure 30.9a). To use the pointer chain as a stack for the recursion of the traversal algorithm, the chain is temporarily reversed to ...➛ C.k ➛ B.j ➛ A.i so that each field in the chain points to its predecessor instead of to its successor (Figure 30.9b).


Each step of the traversal algorithm's recursion must be completed by "popping the stack." After processing any object in the chain (e.g., C), its predecessor (e.g., B) is found by following the reversed pointer chain. The algorithm also needs to know which field of the predecessor was being worked on. To maintain this information, the algorithm must be changed at the earlier stage where it left B to process C. At that stage, the index of the field, j, is copied into the count field of the object table entry of B. The count can be overwritten because the object is being deallocated. But if the size of B exceeds 255 words, then the count field will not be large enough to store every field index. Instead, the allocator is revised to over-allocate by one word any object that is HugeSize (256) words or more and to reserve that extra word for use by the traversal algorithm to store offset.

그림 30-9


To accommodate over-allocation, the allocation routine is revised to accept an additional argument, extraWord, that is either 0 or 1. It is also necessary for the allocator to increment the reference count of the new object's class before storing the class into the header of the new object. (In fact, this must be accomplished even earlier, before calling allocateChunk:, to assure that the class is not deallocated accidentally by some side effect of that subroutine.)

allocate: size extra: extraWord class: classPointer
    "**Preliminary Version**"
    | objectPointer |
    self countUp: classPointer.
        "increment the reference count of the class"
    objectPointer  self allocateChunk: size + extraWord.
        "allocate enough"
    self classBitsOf: objectPointer put: classPointer.
    HeaderSize to: size - 1 do:
        [ :i | self heapChunkOf: objectPointer word: i put: NilPointer].
    "the next statement to correct the SIZE need only be executed if extraWord > 0"
    self sizeBitsOf: objectPointer put: size.
    objectPointer


The actual heap space occupied by an object with at least HugeSize fields is one greater than that stated in its size field, because of the extra word allocated. Therefore, the spaceOccupiedBy: routine must be changed to account for the difference.

spaceOccupiedBy: objectPointer "**Preliminary Version**"
    | size |
    size  self sizeBitsOf: objectPointer.
    size < HugeSize
        ifTrue: [size]
        ifFalse: [size + 1]


The deallocation algorithm must also be revised because deallocated objects have no provision for an extra word not counted in the size field.

deallocate: objectPointer
    | space |
    space  self spaceOccupiedBy: objectPointer.
    self sizeBitsOf: objectPointer put: space.
    self toFreeChunkList: (space min: BigSize) add: objectPointer


The following routine implements the space-efficient traversal algorithm, with A, B, and C of the above example represented by the variables prior, current, and next. To simplify the loop test, the method scans the fields of each chunk in reverse order. Thus the class field is processed last.


Note that the last statement of the method restores the pointer chain to get B.j again pointing to C instead of to A. It is easy to do so when returning to B from processing C, because object pointer of C can simply be stored in the jth field of B. One might think that step unnecessary, because B is being deallocated. However, the same traversal algorithm can be used by a marking collector in which B is not being deallocated.

forAllOtherObjectsAccessibleFrom: objectPointer suchThat: predicate do: action
    | prior current offset size next |
    "compute prior, current, offset, and size to begin processing objectPointer"
    prior  NonPointer.
    current  objectPointer.
    offset  size  self lastPointerOf: objectPointer.
    [true] whileTrue: "for all pointers in all objects traversed"
        [(offset  offset - 1) > 0 "decrement the field index"
            ifTrue: "the class field hasn't been passed yet"
                [next  self heapChunkOf: current word: offset.
                                "one of the pointers"
                    ((self isIntegerObject: next) = = false
                                        and: [predicate value: next])
                        ifTrue: "it's a non-immediate object and it should be processed"
                            ["reverse the pointer chain"
                                self heapChunkOf: current word: offset put: prior.
                                "save the offset either in the count field or in the extra word"
                                size < HugeSize
                                    ifTrue: [self countBitsOf: current put: offset]
                                    ifFalse: [self heapChunkOf: current
                                                    word: size + 1 put: offset].
                                "compute prior, current, offset, and size to begin processing next"
                                prior  current. current  next.
                                offset  size  self lastPointerOf: current]]
            ifFalse:
                ["all pointers have been followed, now perform the action"
                    action value: current.
                    "did we get here from another object?"
                    prior = NonPointer
                        ifTrue: "this was the root object, so we are done"
                            [objectPointer].
                    "restore next, current, and size to resume processing prior"
                    next  current. current  prior.
                    size  self lastPointerOf: current.
                    "restore offset either from the count field or from the extra word"
                    size < HugeSize
                        ifTrue: [offset  self countBitsOf: current]
                        ifFalse: [offset  self heapChunkOf: current word: size + 1].
                    "restore prior from the reversed pointer chain"
                    prior  self heapChunkOf: current word: offset.
                    "restore (unreverse) the pointer chain"
                    self heapChunkOf: current word: offset put: next]]


The machine-language implementation can deal with the procedural arguments either by passing a pair of subroutine addresses to be called indirectly or by expanding the subroutines in line. If the hardware has enough registers, it is possible to keep the variables next, current, prior, size, and offset in registers for additional speed of execution.


A Marking Collector

The job of the marking garbage collector is to mark all accessible objects so that the remaining inaccessible objects can be identified and added to the lists of free chunks. Accessible objects can be found most easily by a recursive search from the "roots of the world," namely, the interpreter's stacks and the table of global variables (the Dictionary named Smalltalk).


The following algorithm is performed on each root object. In the object table entry of the object, set the count field to 1 to mean "marked." Apply the algorithm of this paragraph to each unmarked object referenced by the object.


Note that the above marking algorithm is inherently recursive. In its implementation, the same traversal routine used for reference counting can be used, in either the simple or the space-efficient version. Before marking begins, the count fields of all objects are reset to 0 to mean "unmarked." After marking ends, all unmarked objects are deallocated and the reference counts of all marked objects are recomputed. The routine that performs all the necessary steps is called reclaimInaccessibleObjects.

reclaimInaccessibleObjects
    self zeroReferenceCounts.
    self markAccessibleObjects.
    self rectifyCountsAndDeallocateGarbage


The subroutine that sets the count fields of all objects to 0 is called zeroReferenceCounts. It is superfluous to zero the count field of a free chunk or of a free entry. Nevertheless, the following version zeroes the count field of every entry, because on most computers, it takes less time to zero the first byte of an entry than it takes to test the status of that entry.

zeroReferenceCounts
    0 to: ObjectTableSize - 2 by: 2 do:
        [ :objectPointer |
            self countBitsOf: objectPointer put: 0]


The subroutine markAccessibleObjects invokes the marking algorithm markObjectsAccessibleFrom: for every object in the list rootObjectPointers. Typically, the list rootObjectPointers includes the object pointer of the current process and the object pointer of the global variable dictionary, from which all other accessible objects are referenced directly or indirectly.

markAccessibleObjects
    rootObjectPointers do:
        [ :rootObjectPointer |
            self markObjectsAccessibleFrom: rootObjectPointer]


The marking algorithm markObjectsAccessibleFrom: calls the same traversal routine as the reference-counting collector did. Its predicate succeeds for unmarked objects and it marks them with a count of 1 as a side effect. Its action restores the count field to 1 because the space-efficient version of the traversal routine could have changed that field to any nonzero value as a side effect.

markObjectsAccessibleFrom: rootObjectPointer
    | unmarked |
    self forAllObjectsAccessibleFrom: rootObjectPointer
        suchThat: "the predicate tests for an unmarked object and marks it"
            [ :objectPointer |
                unmarked  (self countBitsOf: objectPointer) = 0.
                unmarked ifTrue: [self countBitsOf: objectPointer put: 1].
                unmarked]
        do: "the action restores the mark to count = 1"
            [ :objectPointer |
                self countBitsOf: objectPointer put: 1]


After the marking algorithm has been executed, every non-free object table entry is examined using the subroutine rectifyCountsAndDeallocateGarbage. If the entry is unmarked, then the entry and its heap chunk are added to the appropriate free lists. If the entry is marked, then the count is decremented by one to unmark it, and the counts of all objects that it references directly are incremented. Note that when a marked object is processed, its count may exceed 1 because objects previously processed may have referenced it. That is why it is unmarked by subtraction instead of by setting its count to 0.


During the examination of object table entries, chunks that were already free before the marking collection began will be encountered. The count field of an already-free chunk is zero just like an unmarked object, so it will be added to a free-chunk list. Doing so would cause a problem if the chunk were already on a free-chunk list. Therefore before the scan begins, all heads of free-chunk lists are reset.


As a final step, the reference count of each root object is incremented to assure that it is not deallocated accidentally.

rectifyCountsAndDeallocateGarbage
    | count |
    "reset heads of free-chunk lists"
    FirstHeapSegment to: LastHeapSegment do: "for every segment"
        [ :segment |
            HeaderSize to: BigSize do: "for every free chunk list"
                [ :size | "reset the list head"
                    self resetFreeChunkList: size inSegment: segment]].
    "rectify counts, and deallocate garbage"
    0 to: ObjectTableSize - 2 by: 2 do: "for every object table entry"
        [ :objectPointer |
            (self freeBitOf: objectPointer) = 0
                ifTrue: "if it is not a free entry"
                    [(count  self countBitsOf: objectPointer) = 0
                        ifTrue: "it is unmarked, so deallocate it"
                            [self deallocate: objectPointer]
                        ifFalse: "it is marked, so rectify reference counts"
                            [count < 128 ifTrue: "subtract 1 to compensate for the mark"
                                    [self countBitsOf: objectPointer put: count - 1].
                                1 to: (self lastPointerOf: objectPointer) - 1 do:
                                    [ :offset | "increment the reference count of each
                                                    pointer"    
                                        self countUp: (self heapChunkOf: objectPointer
                                                            word: offset)]]]].
                                        "be sure the root objects don't disappear"
    rootObjectPointers do:
        [ :rootObjectPointer | self countUp: rootObjectPointer].
    self countBitsOf: NilPointer put: 128


The allocateChunk: routine can now be revised so that it attempts a marking collection if compaction of all segments has failed to yield enough space to satisfy an allocation request.

allocateChunk: size
    | objectPointer |
    objectPointer  self attemptToAllocateChunk: size.
    objectPointer isNil ifFalse: [objectPointer].
    self reclaimInaccessibleObjects. "garbage collect and try again"
    objectPointer  self attemptToAllocateChunk: size.
    objectPointer isNil ifFalse: [objectPointer].
    self outOfMemoryError "give up"


Nonpointer Objects

The object format presented in this chapter is not particularly space efficient, but since its uniformity makes the system software small and simple, the inefficiency can generally be forgiven. There are two classes of object for which the inefficiency is intolerable, namely, character strings and bytecoded methods. There are usually many strings and methods in memory, and when stored one character or one bytecode per word, they are quite wasteful of space.


To store such objects more efficiently, an alternate memory format is used in which the data part of an object contains 8-bit or 16-bit values that are interpreted as unsigned integers rather than as object pointers. Such objects are distinguished by the setting of the pointer-fields bit of the object table entry: when that bit is 1, the data consist of object pointers; when that bit is 0, the data consist of positive 8 or 16-bit integers. When there are an odd number of bytes of data in a nonpointer object, the final byte of the last word is 0 (a slight waste of space), and the odd-length bit of the object table entry, which is normally 0, is set to 1. To support nonpointer objects, the allocator needs two additional parameters, pointerBit and oddBit. In the case of a nonpointer object (pointerBit = 0), the default initial value of the elements is 0 instead of nil. The final version of the allocation routine is shown below.

allocate: size odd: oddBit pointer: pointerBit extra: extraWord class: classPointer
    | objectPointer default |
    self countUp: classPointer.
    objectPointer  self allocateChunk: size + extraWord.
    self oddBitOf: objectPointer put: oddBit.
    self pointerBitOf: objectPointer put: pointerBit.
    self classBitsOf: objectPointer put: classPointer.
    default  pointerBit = 0 ifTrue: [0] ifFalse: [NilPointer].
    HeaderSize to: size - 1 do:
        [ :i | self heapChunkOf: objectPointer word: i put: default].
    self sizeBitsOf: objectPointer put:size.
    objectPointer


The garbage-collecting traversal routines need only process the class field of each nonpointer object, because the data contain no pointers. To make this happen, the routine lastPointerOf: is changed as follows:

lastPointerOf: objectPointer "**Preliminary Version**"
    (self pointerBitOf: objectPointer) = 0
        ifTrue:
            [HeaderSize]
        ifFalse:
            [self sizeBitsOf: objectPointer]


The value of lastPointerOf: is never as large as 256 for a nonpointer object, so a nonpointer object never needs to be over-allocated. Therefore, spaceOccupiedBy: is revised again as follows:

spaceOccupiedBy: objectPointer
    | size |
    size  self sizeBitsOf: objectPointer.
    (size < HugeSize or: [(self pointerBitOf: objectPointer) = 0])
        ifTrue: [size]
        ifFalse: [size + 1]


CompiledMethods

A CompiledMethod is an anomaly for the memory manager because its data are a mixture of 16-bit pointers and 8-bit unsigned integers. The only change needed to support CompiledMethods is to add to lastPointerOf: a computation similar to that in the bytecode interpreter's routine bytecodeIndexOf:. MethodClass is the object table index of CompiledMethod.

lastPointerOf: objectPointer
    | methodHeader |
    (self pointerBitOf: objectPointer) = 0
        ifTrue:
            [HeaderSize]
        ifFalse:
            [(self classBitsOf: objectPointer) = MethodClass
                ifTrue: [methodHeader  self heapChunkOf: objectPointer
                                                word: HeaderSize.
                            HeaderSize + 1 + ((methodHeader bitAnd: 126)
                                                        bitShift: -1)]
                ifFalse: [self sizeBitsOf: objectPointer]]


Interface to the Bytecode Interpreter

The final step in the implementation of the object memory is to provide the interface routines required by the interpreter. Note that fetchClassOf: objectPointer returns IntegerClass (the object table index of SmallInteger) if its argument is an immediate integer.

object pointer access
fetchPointer: fieldIndex ofObject: objectPointer
    self heapChunkOf: objectPointer word: HeaderSize + fieldIndex
storePointer: fieldIndex ofObject: objectPointer withValue: valuePointer
    | chunkIndex |
    chunkIndex  HeaderSize + fieldIndex.
    self countUp: valuePointer.
    self countDown: (self heapChunkOf: objectPointer word: chunkIndex).
    self heapChunkOf: objectPointer word: chunkIndex put: valuePointer

word access
fetchWord: wordIndex ofObject: objectPointer
    self heapChunkOf: objectPointer word: HeaderSize + wordIndex
storeWord: wordIndex ofObject: objectPointer withValue: valueWord
    self heapChunkOf: objectPointer word: HeaderSize + wordIndex
        put: valueWord

byte access
fetchByte: byteIndex ofObject: objectPointer
    self heapChunkOf: objectPointer byte: (HeaderSize * 2 + byteIndex)
storeByte: byteIndex ofObject: objectPointer withValue: valueByte
    self heapChunkOf: objectPointer
        byte: (HeaderSize * 2 + byteIndex)
        put: valueByte

regerence counting
increaseReferencesTo: objectPointer
    self countUp: objectPointer
decreaseReferencesTo: objectPointer
    self countDown: objectPointer

class pointer access
fetchClassOf: objectPointer
    (self isInteger Object: objectPointer)
        ifTrue: [IntegerClass]
        ifFalse: [self classBitsOf: objectPointer]

length access
fetchWordLengthOf: objectPointer
    (self sizeBitsOf: objectPointer) - HeaderSize
fetchByteLengthOf: objectPointer
    (self loadWordLengthOf: objectPointer) * 2 - (self oddBitOf: objectPointer)

object creation
instantiateClass: classPointer withPointers: length
    | size extra |
    size  HeaderSize + length.
    extra  size < HugeSize ifTrue: [0] ifFalse: [1].
    self allocate: size odd: 0 pointer: 1 extra: extra class: classPointer
instantiateClass: classPointer withWords: length
    | size |
    size  HeaderSize + length.
    self allocate: size odd: 0 pointer: 0 extra: 0 class: classPointer
instantiateClass: classPointer withBytes: length
    | size |
    size  HeaderSize + ((length + 1)/2).
    self allocate: size odd: length \\ 2 pointer: 0 extra: 0 class: classPointer

instance enumeration
initialInstanceOf: classPointer
    0 to: ObjectTableSize - 2 by: 2 do:
        [ :pointer |
            (self freeBitOf: pointer) = 0
                ifTrue: [(self fetchClassOf: pointer) = classPointer
                    ifTrue: [pointer]]].
    NilPointer
instanceAfter: objectPointer
    | classPointer |
    objectPointer to: ObjectTableSize - 2 by: 2 do:
        [ :pointer |
            (self freeBitOf: pointer) = 0
                ifTrue: [(self fetchClassOf: pointer) = classPointer
                    ifTrue: [pointer]]].
    NilPointer

pointer swapping
swapPointersOf: firstPointer and: secondPointer
    | firstSegment firstLocation firstPointer firstOdd |
    firstSegment  self segmentBitsOf: firstPointer.
    firstLocation  self locationBitsOf: firstPointer.
    firstPointer  self pointerBitOf: firstPointer.
    firstOdd  self oddBitOf: firstPointer.
    self segmentBitsOf: firstPointer put: (self segmentBitsOf: secondPointer).
    self locationBitsOf: firstPointer put: (self locationBitsOf: secondPointer).
    self pointerBitOf: firstPointer put: (self pointerBitOf: secondPointer).
    self oddBitOf: firstPointer put: (self oddBitOf: secondPointer).
    self segmentBitsOf: secondPointer put: firstSegment.
    self locationBitsOf: secondPointer put: firstLocation.
    self pointerBitOf: secondPointer put: firstPointer.
    self oddBitOf: secondPointer put: firstOdd

integer access
integerValueOf: objectPointer
    objectPointer / 2
integerObjectOf: value
    (value bitShift: 1) + 1
isIntegerObject: objectPointer
    (objectPointer bitAnd: 1) = 1
isIntegerValue: valueWord
    valueWord < = -16384 and: [valueWord > 16384]


Notes