Scheme is a latently-typed language; this means that the system cannot, in general, determine the type of a given expression at compile time. Types only become apparent at run time. Variables do not have fixed types; a variable may hold a pair at one point, an integer at the next, and a thousand-element vector later. Instead, values, not variables, have fixed types.
In order to implement standard Scheme functions like pair?
and
string?
and provide garbage collection, the representation of
every value must contain enough information to accurately determine its
type at run time. Often, Scheme systems also use this information to
determine whether a program has attempted to apply an operation to an
inappropriately typed value (such as taking the car
of a string).
Because variables, pairs, and vectors may hold values of any type, Scheme implementations use a uniform representation for values -- a single type large enough to hold either a complete value or a pointer to a complete value, along with the necessary typing information.
The following sections will present a simple typing system, and then make some refinements to correct its major weaknesses. However, this is not a description of the system Guile actually uses. It is only an illustration of the issues Guile's system must address. We provide all the information one needs to work with Guile's data in section How Guile does it.
The simplest way to meet the above requirements in C would be to
represent each value as a pointer to a structure containing a type
indicator, followed by a union carrying the real value. Assuming that
SCM
is the name of our universal type, we can write:
enum type { integer, pair, string, vector, ... }; typedef struct value *SCM; struct value { enum type type; union { int integer; struct { SCM car, cdr; } pair; struct { int length; char *elts; } string; struct { int length; SCM *elts; } vector; ... } value; };
with the ellipses replaced with code for the remaining Scheme types.
This representation is sufficient to implement all of Scheme's
semantics. If x is an SCM
value:
x->type == integer
.
x->value.integer
.
x->type == vector
.
x->value.vector.elts[0]
to refer to its first element.
x->value.pair.car
to extract its car.
Unfortunately, the above representation has a serious disadvantage. In
order to return an integer, an expression must allocate a struct
value
, initialize it to represent that integer, and return a pointer to
it. Furthermore, fetching an integer's value requires a memory
reference, which is much slower than a register reference on most
processors. Since integers are extremely common, this representation is
too costly, in both time and space. Integers should be very cheap to
create and manipulate.
One possible solution comes from the observation that, on many
architectures, structures must be aligned on a four-byte boundary.
(Whether or not the machine actually requires it, we can write our own
allocator for struct value
objects that assures this is true.)
In this case, the lower two bits of the structure's address are known to
be zero.
This gives us the room we need to provide an improved representation for integers. We make the following rules:
SCM
value are zero, then the SCM
value is a pointer to a struct value
, and everything proceeds as
before.
SCM
value represents an integer, whose value
appears in its upper bits.
Here is C code implementing this convention:
enum type { pair, string, vector, ... }; typedef struct value *SCM; struct value { enum type type; union { struct { SCM car, cdr; } pair; struct { int length; char *elts; } string; struct { int length; SCM *elts; } vector; ... } value; }; #define POINTER_P(x) (((int) (x) & 3) == 0) #define INTEGER_P(x) (! POINTER_P (x)) #define GET_INTEGER(x) ((int) (x) >> 2) #define MAKE_INTEGER(x) ((SCM) (((x) << 2) | 1))
Notice that integer
no longer appears as an element of enum
type
, and the union has lost its integer
member. Instead, we
use the POINTER_P
and INTEGER_P
macros to make a coarse
classification of values into integers and non-integers, and do further
type testing as before.
Here's how we would answer the questions posed above (again, assume
x is an SCM
value):
INTEGER_P (x)
.
GET_INTEGER (x)
.
POINTER_P (x) && x->type == vector
Given the new representation, we must make sure x is truly a
pointer before we dereference it to determine its complete type.
x->value.vector.elts[0]
to refer to its first element, as
before.
x->value.pair.car
to extract its car, just as before.
This representation allows us to operate more efficiently on integers than the first. For example, if x and y are known to be integers, we can compute their sum as follows:
MAKE_INTEGER (GET_INTEGER (x) + GET_INTEGER (y))
Now, integer math requires no allocation or memory references. Most
real Scheme systems actually use an even more efficient representation,
but this essay isn't about bit-twiddling. (Hint: what if pointers had
01
in their least significant bits, and integers had 00
?)
However, there is yet another issue to confront. Most Scheme heaps
contain more pairs than any other type of object; Jonathan Rees says
that pairs occupy 45% of the heap in his Scheme implementation, Scheme
48. However, our representation above spends three SCM
-sized
words per pair -- one for the type, and two for the CAR and
CDR. Is there any way to represent pairs using only two words?
Let us refine the convention we established earlier. Let us assert that:
SCM
value are #b00
, then
it is a pointer, as before.
#b01
, then the upper bits are an
integer. This is a bit more restrictive than before.
#b10
, then the value, with the bottom
two bits masked out, is the address of a pair.
Here is the new C code:
enum type { string, vector, ... }; typedef struct value *SCM; struct value { enum type type; union { struct { int length; char *elts; } string; struct { int length; SCM *elts; } vector; ... } value; }; struct pair { SCM car, cdr; }; #define POINTER_P(x) (((int) (x) & 3) == 0) #define INTEGER_P(x) (((int) (x) & 3) == 1) #define GET_INTEGER(x) ((int) (x) >> 2) #define MAKE_INTEGER(x) ((SCM) (((x) << 2) | 1)) #define PAIR_P(x) (((int) (x) & 3) == 2) #define GET_PAIR(x) ((struct pair *) ((int) (x) & ~3))
Notice that enum type
and struct value
now only contain
provisions for vectors and strings; both integers and pairs have become
special cases. The code above also assumes that an int
is large
enough to hold a pointer, which isn't generally true.
Our list of examples is now as follows:
INTEGER_P
(x)
; this is as before.
GET_INTEGER (x)
, as
before.
POINTER_P (x) && x->type == vector
We must still make sure that x is a pointer to a struct
value
before dereferencing it to find its type.
x->value.vector.elts[0]
to refer to its first element, as
before.
PAIR_P (x)
to determine if x is a
pair, and then write GET_PAIR (x)->car
to refer to its
car.
This change in representation reduces our heap size by 15%. It also
makes it cheaper to decide if a value is a pair, because no memory
references are necessary; it suffices to check the bottom two bits of
the SCM
value. This may be significant when traversing lists, a
common activity in a Scheme system.
Again, most real Scheme systems use a slighty different implementation;
for example, if GET_PAIR subtracts off the low bits of x
, instead
of masking them off, the optimizer will often be able to combine that
subtraction with the addition of the offset of the structure member we
are referencing, making a modified pointer as fast to use as an
unmodified pointer.
We originally started with a very simple typing system -- each object
has a field that indicates its type. Then, for the sake of efficiency
in both time and space, we moved some of the typing information directly
into the SCM
value, and left the rest in the struct value
.
Guile itself employs a more complex hierarchy, storing finer and finer
gradations of type information in different places, depending on the
object's coarser type.
In the author's opinion, Guile could be simplified greatly without significant loss of efficiency, but the simplified system would still be more complex than what we've presented above.
Here we present the specifics of how Guile represents its data. We don't go into complete detail; an exhaustive description of Guile's system would be boring, and we do not wish to encourage people to write code which depends on its details anyway. We do, however, present everything one need know to use Guile's data.
Any code which operates on Guile datatypes must #include
the
header file <libguile.h>
. This file contains a definition for
the SCM
typedef (Guile's universal type, as in the examples
above), and definitions and declarations for a host of macros and
functions that operate on SCM
values.
All identifiers declared by <libguile.h>
begin with scm_
or SCM_
.
The functions described here generally check the types of their
SCM
arguments, and signal an error if their arguments are of an
inappropriate type. Macros generally do not, unless that is their
specified purpose. You must verify their argument types beforehand, as
necessary.
Macros and functions that return a boolean value have names ending in
P
or _p
(for "predicate"). Those that return a negated
boolean value have names starting with SCM_N
. For example,
SCM_IMP (x)
is a predicate which returns non-zero iff
x is an immediate value (an IM
). SCM_NCONSP
(x)
is a predicate which returns non-zero iff x is
not a pair object (a CONS
).
Aside from the latent typing, the major source of constraints on a Scheme implementation's data representation is the garbage collector. The collector must be able to traverse every live object in the heap, to determine which objects are not live.
There are many ways to implement this, but Guile uses an algorithm called mark and sweep. The collector scans the system's global variables and the local variables on the stack to determine which objects are immediately accessible by the C code. It then scans those objects to find the objects they point to, et cetera. The collector sets a mark bit on each object it finds, so each object is traversed only once. This process is called tracing.
When the collector can find no unmarked objects pointed to by marked objects, it assumes that any objects that are still unmarked will never be used by the program (since there is no path of dereferences from any global or local variable that reaches them) and deallocates them.
In the above paragraphs, we did not specify how the garbage collector finds the global and local variables; as usual, there are many different approaches. Frequently, the programmer must maintain a list of pointers to all global variables that refer to the heap, and another list (adjusted upon entry to and exit from each function) of local variables, for the collector's benefit.
The list of global variables is usually not too difficult to maintain, since global variables are relatively rare. However, an explicitly maintained list of local variables (in the author's personal experience) is a nightmare to maintain. Thus, Guile uses a technique called conservative garbage collection, to make the local variable list unnecessary.
The trick to conservative collection is to treat the stack as an ordinary range of memory, and assume that every word on the stack is a pointer into the heap. Thus, the collector marks all objects whose addresses appear anywhere in the stack, without knowing for sure how that word is meant to be interpreted.
Obviously, such a system will occasionally retain objects that are actually garbage, and should be freed. In practice, this is not a problem. The alternative, an explicitly maintained list of local variable addresses, is effectively much less reliable, due to programmer error.
To accomodate this technique, data must be represented so that the collector can accurately determine whether a given stack word is a pointer or not. Guile does this as follows:
Thus, given any random word w fetched from the stack, Guile's garbage collector can consult the table to see if w falls within a known heap segment, and check w's alignment. If both tests pass, the collector knows that w is a valid pointer to a cell, intentional or not, and proceeds to trace the cell.
Note that heap segments do not contain all the data Guile uses; cells for objects like vectors and strings contain pointers to other memory areas. However, since those pointers are internal, and not shared among many pieces of code, it is enough for the collector to find the cell, and then use the cell's type to find more pointers to trace.
Guile classifies Scheme objects into two kinds: those that fit entirely
within an SCM
, and those that require heap storage.
The former class are called immediates. The class of immediates includes small integers, characters, boolean values, the empty list, the mysterious end-of-file object, and some others.
The remaining types are called, not suprisingly, non-immediates. They include pairs, procedures, strings, vectors, and all other data types in Guile.
SCM_IMP
, above.
You must use this macro before calling a finer-grained predicate to determine x's type. For example, to see if x is a pair, you must write:
SCM_NIMP (x) && SCM_CONSP (x)
This is because Guile stores typing information for non-immediate values
in their cells, rather than in the SCM
value itself; thus, you
must determine whether x refers to a cell before looking inside
it.
This is somewhat of a pity, because it means that the programmer needs
to know which types Guile implements as immediates vs. non-immediates.
There are (possibly better) representations in which SCM_CONSP
can be self-sufficient. The immediate type predicates do not suffer
from this weakness.
The following datatypes are immediate values; that is, they fit entirely
within an SCM
value. The SCM_IMP
and SCM_NIMP
macros will distinguish these from non-immediates; see section Immediates vs. Non-immediates for an explanation of the distinction.
Note that the type predicates for immediate values work correctly on any
SCM
value; you do not need to call SCM_IMP
first, to
establish that a value is immediate. This differs from the
non-immediate type predicates, which work correctly only on
non-immediate values; you must be sure the value is SCM_NIMP
before applying them.
Here are functions for operating on small integers, that fit within an
SCM
. Such integers are called immediate numbers, or
INUMs. In general, INUMs occupy all but two bits of an
SCM
.
Bignums and floating-point numbers are non-immediate objects, and have their own, separate accessors. The functions here will not work on them. This is not as much of a problem as you might think, however, because the system never constructs bignums that could fit in an INUM, and never uses floating point values for exact integers.
SCM
.
This function does not check for overflow.
Here are functions for operating on characters.
x
as a C character. If x is not a
Scheme character, the result is undefined.
Here are functions and macros for operating on booleans.
#f
is true, this amounts to comparing x to
#f
; hence the name.
The immediate values that are neither small integers, characters, nor booleans are all unique values -- that is, datatypes with only one instance.
'()
.
This is sort of a weirdly literal way to take things, but the standard read-eval-print loop prints nothing when the expression returns this value, so it's not a bad idea to return this when you can't think of anything else helpful.
For example, when you write a C function that is callable from Scheme
and which takes optional arguments, the interpreter passes
SCM_UNDEFINED
for any arguments you did not receive.
We also use this to mark unbound variables.
SCM_UNDEFINED
. Apply this to a
symbol's value to see if it has a binding as a global variable.
A non-immediate datatype is one which lives in the heap, either because
it cannot fit entirely within a SCM
word, or because it denotes a
specific storage location (in the nomenclature of the Revised^4 Report
on Scheme).
The SCM_IMP
and SCM_NIMP
macros will distinguish these
from immediates; see section Immediates vs. Non-immediates.
Given a cell, Guile distinguishes between pairs and other non-immediate types by storing special tag values in a non-pair cell's car, that cannot appear in normal pairs. A cell with a non-tag value in its car is an ordinary pair. The type of a cell with a tag in its car depends on the tag; the non-immediate type predicates test this value. If a tag value appears elsewhere (in a vector, for example), the heap may become corrupted.
As mentioned in section Garbage Collection, all non-immediate objects
start with a cell, or a pair of words. Furthermore, all type
information that distinguishes one kind of non-immediate from another is
stored in the cell. The type information in the SCM
value
indicates only that the object is a non-immediate; all finer
distinctions require one to examine the cell itself, usually with the
appropriate type predicate macro.
The type predicates for non-immediate objects generally assume that
their argument is a non-immediate value. Thus, you must be sure that a
value is SCM_NIMP
first before passing it to a non-immediate type
predicate. Thus, the idiom for testing whether a value is a cell or not
is:
SCM_NIMP (x) && SCM_CONSP (x)
Pairs are the essential building block of list structure in Scheme. A pair object has two fields, called the car and the cdr.
It is conventional for a pair's CAR to contain an element of a
list, and the CDR to point to the next pair in the list, or to
contain SCM_EOL
, indicating the end of the list. Thus, a set of
pairs chained through their CDRs constitutes a singly-linked list.
Scheme and libguile define many functions which operate on lists
constructed in this fashion, so although lists chained through the
CARs of pairs will work fine too, they may be less convenient to
manipulate, and receive less support from the community.
Guile implements pairs by mapping the CAR and CDR of a pair directly into the two words of the cell.
This is the most primitive way to allocate a cell; it is quite fast.
The CAR of the cell initially tags it as a "free cell". If the caller intends to use it as an ordinary cons, she must store ordinary SCM values in its CAR and CDR.
If the caller intends to use it as a header for some other type, she
must store an appropriate magic value in the cell's CAR, to mark
it as a member of that type, and store whatever value in the CDR
that type expects. You should generally not do this, unless you are
implementing a new datatype, and thoroughly understand the code in
<libguile/tags.h>
.
The macros below perform no typechecking. The results are undefined if cell is an immediate. However, since all non-immediate Guile objects are constructed from cells, and these macros simply return the first element of a cell, they actually can be useful on datatypes other than pairs. (Of course, it is not very modular to use them outside of the code which implements that datatype.)
Vectors, strings, and symbols have some properties in common. They all
have a length, and they all have an array of elements. In the case of a
vector, the elements are SCM
values; in the case of a string or
symbol, the elements are characters.
All these types store their length (along with some tagging bits) in the
CAR of their header cell, and store a pointer to the elements in
their CDR. Thus, the SCM_CAR
and SCM_CDR
macros
are (somewhat) meaningful when applied to these datatypes.
There are also a few magic values stuffed into memory before a symbol's characters, but you don't want to know about those. What cruft!
Guile provides two kinds of procedures: closures, which are the
result of evaluating a lambda
expression, and subrs, which
are C functions packaged up as Scheme objects, to make them available to
Scheme programmers.
(There are actually other sorts of procedures: compiled closures, and continuations; see the source code for details about them.)
SCM_BOOL_T
iff x is a Scheme procedure object, of
any sort. Otherwise, return SCM_BOOL_F
.
[FIXME: this needs to be further subbed, but texinfo has no subsubsub]
A closure is a procedure object, generated as the value of a
lambda
expression in Scheme. The representation of a closure is
straightforward -- it contains a pointer to the code of the lambda
expression from which it was created, and a pointer to the environment
it closes over.
In Guile, each closure also has a property list, allowing the system to store information about the closure. I'm not sure what this is used for at the moment -- the debugger, maybe?
This function should probably only be used internally by the interpreter, since the representation of the code is intimately connected with the interpreter's implementation.
This function should probably only be used internally by the interpreter, since the representation of the environment is intimately connected with the interpreter's implementation.
[FIXME: this needs to be further subbed, but texinfo has no subsubsub]
A subr is a pointer to a C function, packaged up as a Scheme object to make it callable by Scheme code. In addition to the function pointer, the subr also contains a pointer to the name of the function, and information about the number of arguments accepted by the C fuction, for the sake of error checking.
There is no single type predicate macro that recognizes subrs, as
distinct from other kinds of procedures. The closest thing is
scm_procedure_p
; see section Procedures.
The subr object accepts req required arguments, opt optional
arguments, and a rest argument iff rest is non-zero. The C
function function should accept req + opt
arguments, or req + opt + 1
arguments if rest
is non-zero.
When a subr object is applied, it must be applied to at least req
arguments, or else Guile signals an error. function receives the
subr's first req arguments as its first req arguments. If
there are fewer than opt arguments remaining, then function
receives the value SCM_UNDEFINED
for any missing optional
arguments. If rst is non-zero, then any arguments after the first
req + opt
are packaged up as a list as passed as
function's last argument.
Note that subrs can actually only accept a predefined set of
combinations of required, optional, and rest arguments. For example, a
subr can take one required argument, or one required and one optional
argument, but a subr can't take one required and two optional arguments.
It's bizarre, but that's the way the interpreter was written. If the
arguments to scm_make_gsubr
do not fit one of the predefined
patterns, then scm_make_gsubr
will return a compiled closure
object instead of a subr object.
Haven't written this yet, 'cos I don't understand ports yet.
Every function visible at the Scheme level should aggressively check the types of its arguments, to avoid misinterpreting a value, and perhaps causing a segmentation fault. Guile provides some macros to make this easier.
If position is a string, SCM_ASSERT
raises a
"miscellaneous" error whose message is that string.
Otherwise, position should be one of the values defined below.
SCM_ASSERT
, SCM_ARGn
claims that
obj has the wrong type for the n'th argument of subr.
The only way to complain about the type of an argument after the fifth
is to use SCM_ARGn
, defined below, which doesn't specify which
argument is wrong. You could pass your own error message to
SCM_ASSERT
as the position, but then the error signalled is
a "miscellaneous" error, not a "wrong type argument" error. This
seems kludgy to me.
Interestingly, the message is attributed to the function named by obj, not subr, so obj must be a Scheme string object naming the function. Usually, Guile catches these errors before ever invoking the subr, so we don't run into these problems.
Smobs are Guile's mechanism for adding new non-immediate types to
the system.(1) To define a new smob type, the programmer provides Guile with
some essential information about the type -- how to print it, how to
garbage collect it, and so on -- and Guile returns a fresh type tag for
use in the CAR of new cells. The programmer can then use
scm_make_gsubr
to make a set of C functions that create and
operate on these objects visible to Scheme code.
(You can find a complete version of the example code used in this
section in the Guile distribution, in `doc/example-smob'. That
directory includes a makefile and a suitable main
function, so
you can build a complete interactive Guile shell, extended with the
datatypes described here.)
To define a new type, the programmer must fill in an scm_smobfuns
structure with functions to manage instances of the type. Here is the
definition of the structure:
typedef struct scm_smobfuns { SCM (*mark) (SCM obj); scm_sizet (*free) (SCM obj); int (*print) (SCM obj, SCM port, scm_print_state *pstate); SCM (*equalp) (SCM a, SCM b); } scm_smobfuns;
mark
free
print
display
or write
. The function should
write a printed representation of exp on port, in accordance
with the parameters in pstate. (For more information on print
states, see section Ports.)
equalp
equal?
function to compare two instances
of the same smob type, Guile calls this function. It should return
SCM_BOOL_T
if a and b should be considered
equal?
, or SCM_BOOL_F
otherwise. If equalp
is
zero, equal?
will assume that two instances of this type are
never equal?
unless they are eq?
.
Once you have built a scm_smobfuns
structure, you can call the
scm_newsmob
function to add the type to the system.
For example, here is how one might declare and register a new type representing eight-bit grayscale images:
#include <libguile.h> long image_tag; scm_smobfuns image_funs = { mark_image, free_image, print_image, 0 }; void init_image_type () { image_tag = scm_newsmob (&image_funs); }
Like other non-immediate types, smobs start with a cell whose CAR
contains typing information, and whose cdr
is free for any use.
To create an instance of a smob type, you must allocate a fresh cell, by
calling SCM_NEWCELL
, and store the tag returned by
scm_smobfuns
in its car.
Guile provides the following functions for managing memory, which are often helpful when implementing smobs:
malloc
, and return a
pointer to them.
If there is not enough memory available, invoke the garbage collector, and try once more. If there is still not enough, signal an error, reporting that we could not allocate what.
This function also helps maintain statistics about the size of the heap.
realloc
. Return a
pointer to the new block.
If there is not enough memory available, invoke the garbage collector, and try once more. If there is still not enough, signal an error, reporting that we could not allocate what.
The value olen should be the old size of the block of memory at addr; it is only used for keeping statistics on the size of the heap.
free
. If
addr is zero, signal an error, complaining of an attempt to free
something that is already free.
This does no record-keeping; instead, the smob's free
function
must take care of that.
This function isn't usually sufficiently different from the usual
free
function to be worth using.
Continuing the above example, if the global variable image_tag
contains a tag returned by scm_newsmob
, here is how we could
construct a smob whose CDR contains a pointer to a freshly
allocated struct image
:
struct image { int width, height; char *pixels; /* The name of this image */ SCM name; /* A function to call when this image is modified, e.g., to update the screen, or SCM_BOOL_F if no action necessary */ SCM update_func; }; SCM make_image (SCM name, SCM s_width, SCM s_height) { struct image *image; SCM image_smob; int width, height; SCM_ASSERT (SCM_NIMP (name) && SCM_STRINGP (name), name, SCM_ARG1, "make-image"); SCM_ASSERT (SCM_INUMP (s_width), s_width, SCM_ARG2, "make-image"); SCM_ASSERT (SCM_INUMP (s_height), s_height, SCM_ARG3, "make-image"); width = SCM_INUM (s_width); height = SCM_INUM (s_height); image = (struct image *) scm_must_malloc (sizeof (struct image), "image"); image->width = width; image->height = height; image->pixels = scm_must_malloc (width * height, "image pixels"); image->name = name; image->update_func = SCM_BOOL_F; SCM_NEWCELL (image_smob); SCM_SETCDR (image_smob, image); SCM_SETCAR (image_smob, image_tag); return image_smob; }
Functions that operate on smobs should aggressively check the types of
their arguments, to avoid misinterpreting some other datatype as a smob,
and perhaps causing a segmentation fault. Fortunately, this is pretty
simple to do. The function need only verify that its argument is a
non-immediate, whose CAR is the type tag returned by
scm_newsmob
.
For example, here is a simple function that operates on an image smob,
and checks the type of its argument. We also present an expanded
version of the init_image_type
function, to make
clear_image
and the image constructor function make_image
visible to Scheme code.
SCM clear_image (SCM image_smob) { int area; struct image *image; SCM_ASSERT ((SCM_NIMP (image_smob) && SCM_CAR (image_smob) == image_tag), image_smob, SCM_ARG1, "clear-image"); image = (struct image *) SCM_CDR (image_smob); area = image->width * image->height; memset (image->pixels, 0, area); /* Invoke the image's update function. */ if (image->update_func != SCM_BOOL_F) scm_apply (image->update_func, SCM_EOL, SCM_EOL); return SCM_UNSPECIFIED; } void init_image_type () { image_tag = scm_newsmob (&image_funs); scm_make_gsubr ("make-image", 3, 0, 0, make_image); scm_make_gsubr ("clear-image", 1, 0, 0, clear_image); }
Note that checking types is a little more complicated during garbage
collection; see the description of SCM_GCTYP16
in section Garbage Collecting Smobs.
Once a smob has been released to the tender mercies of the Scheme
system, it must be prepared to survive garbage collection. Guile calls
the mark
and free
functions of the scm_smobfuns
structure to manage this.
As described before (see section Garbage Collection), every object in the Scheme system has a mark bit, which the garbage collector uses to tell live objects from dead ones. When collection starts, every object's mark bit is clear. The collector traces pointers through the heap, starting from objects known to be live, and sets the mark bit on each object it encounters. When it can find no more unmarked objects, the collector walks all objects, live and dead, frees those whose mark bits are still clear, and clears the mark bit on the others.
The two main portions of the collection are called the mark phase, during which the collector marks live objects, and the sweep phase, during which the collector frees all unmarked objects.
The mark bit of a smob lives in its CAR, along with the smob's type
tag. When the collector encounters a smob, it sets the smob's mark bit,
and uses the smob's type tag to find the appropriate mark
function for that smob: the one listed in that smob's
scm_smobfuns
structure. It then calls the mark
function,
passing it the smob as its only argument.
The mark
function is responsible for marking any other Scheme
objects the smob refers to. If it does not do so, the objects' mark
bits will still be clear when the collector begins to sweep, and the
collector will free them. If this occurs, it will probably break, or at
least confuse, any code operating on the smob; the smob's SCM
values will have become dangling references.
To mark an arbitrary Scheme object, the mark
function may call
this function:
Thus, here is how we might write the mark
function for the image
smob type discussed above:
SCM mark_image (SCM image_smob) { /* Mark the image's name and update function. */ struct image *image = (struct image *) SCM_CDR (image_smob); scm_gc_mark (image->name); scm_gc_mark (image->update_func); return SCM_BOOL_F; }
Note that, even though the image's update_func
could be an
arbitrarily complex structure (representing a procedure and any values
enclosed in its environment), scm_gc_mark
will recurse as
necessary to mark all its components. Because scm_gc_mark
sets
an object's mark bit before it recurses, it is not confused by
circular structures.
As an optimization, the collector will mark whatever value is returned
by the mark
function; this helps limit depth of recursion during
the mark phase. Thus, the code above could also be written as:
SCM mark_image (SCM image_smob) { /* Mark the image's name and update function. */ struct image *image = (struct image *) SCM_CDR (image_smob); scm_gc_mark (image->name); return image->update_func; }
Finally, when the collector encounters an unmarked smob during the sweep
phase, it uses the smob's tag to find the appropriate free
function for the smob. It then calls the function, passing it the smob
as its only argument.
The free
function must release any resources used by the smob.
However, it need not free objects managed by the collector; the
collector will take care of them. The return type of the free
function should be scm_sizet
, an unsigned integral type; the
free
function should return the number of bytes released, to help
the collector maintain statistics on the size of the heap.
Here is how we might write the free
function for the image smob
type:
scm_sizet free_image (SCM image_smob) { struct image *image = (struct image *) SCM_CDR (image_smob); scm_sizet size = image->width * image->height + sizeof (*image); free (image->pixels); free (image); return size; }
During the sweep phase, the garbage collector will clear the mark bits on all live objects. The code which implements a smob need not do this itself.
There is no way for smob code to be notified when collection is complete.
Note that, since a smob's mark bit lives in its CAR, along with the smob's type tag, the technique for checking the type of a smob described in section Typechecking will not necessarily work during GC. If you need to find out whether a given object is a particular smob type during GC, use the following macro:
Use this macro instead of SCM_CAR
to check the type of a smob
during GC. Usually, only code called by the smob's mark
function
need worry about this.
It is usually a good idea to minimize the amount of processing done
during garbage collection; keep mark
and free
functions
very simple. Since collections occur at unpredictable times, it is easy
for any unusual activity to interfere with normal code.
It is often useful to define very simple smob types -- smobs which have
no data to mark, other than the cell itself, or smobs whose CDR is
simply an ordinary Scheme object, to be marked recursively. Guile
provides some functions to handle these common cases; you can use these
functions as your smob type's mark
function, if your smob's
structure is simple enough.
If the smob refers to no other Scheme objects, then no action is necessary; the garbage collector has already marked the smob cell itself. In that case, you can use zero as your mark function.
scm_markcdr
as their marking functions, and
refer to no heap storage, including memory managed by malloc
,
other than the smob's header cell.
Here is the complete text of the implementation of the image datatype,
as presented in the sections above. We also provide a definition for
the smob's print
function, and make some objects and functions
static, to clarify exactly what the surrounding code is using.
As mentioned above, you can find this code in the Guile distribution, in
`doc/example-smob'. That directory includes a makefile and a
suitable main
function, so you can build a complete interactive
Guile shell, extended with the datatypes described here.)
/* file "image-type.c" */ #include <stdlib.h> #include <libguile.h> static long image_tag; struct image { int width, height; char *pixels; /* The name of this image */ SCM name; /* A function to call when this image is modified, e.g., to update the screen, or SCM_BOOL_F if no action necessary */ SCM update_func; }; static SCM make_image (SCM name, SCM s_width, SCM s_height) { struct image *image; SCM image_smob; int width, height; SCM_ASSERT (SCM_NIMP (name) && SCM_STRINGP (name), name, SCM_ARG1, "make-image"); SCM_ASSERT (SCM_INUMP (s_width), s_width, SCM_ARG2, "make-image"); SCM_ASSERT (SCM_INUMP (s_height), s_height, SCM_ARG3, "make-image"); width = SCM_INUM (s_width); height = SCM_INUM (s_height); image = (struct image *) scm_must_malloc (sizeof (struct image), "image"); image->width = width; image->height = height; image->pixels = scm_must_malloc (width * height, "image pixels"); image->name = name; image->update_func = SCM_BOOL_F; SCM_NEWCELL (image_smob); SCM_SETCDR (image_smob, image); SCM_SETCAR (image_smob, image_tag); return image_smob; } static SCM clear_image (SCM image_smob) { int area; struct image *image; SCM_ASSERT ((SCM_NIMP (image_smob) && SCM_CAR (image_smob) == image_tag), image_smob, SCM_ARG1, "clear-image"); image = (struct image *) SCM_CDR (image_smob); area = image->width * image->height; memset (image->pixels, 0, area); /* Invoke the image's update function. */ if (image->update_func != SCM_BOOL_F) scm_apply (image->update_func, SCM_EOL, SCM_EOL); return SCM_UNSPECIFIED; } static SCM mark_image (SCM image_smob) { struct image *image = (struct image *) SCM_CDR (image_smob); scm_gc_mark (image->name); return image->update_func; } static scm_sizet free_image (SCM image_smob) { struct image *image = (struct image *) SCM_CDR (image_smob); scm_sizet size = image->width * image->height + sizeof (struct image); free (image->pixels); free (image); return size; } static int print_image (SCM image_smob, SCM port, scm_print_state *pstate) { struct image *image = (struct image *) SCM_CDR (image_smob); scm_puts ("#<image ", port); scm_display (image->name, port); scm_puts (">", port); /* non-zero means success */ return 1; } static scm_smobfuns image_funs = { mark_image, free_image, print_image, 0 }; void init_image_type () { image_tag = scm_newsmob (&image_funs); scm_make_gsubr ("clear-image", 1, 0, 0, clear_image); scm_make_gsubr ("make-image", 3, 0, 0, make_image); }
Here is a sample build and interaction with the code from the `example-smob' directory, on the author's machine:
zwingli:example-smob$ make CC=gcc gcc `guile-config compile` -c image-type.c -o image-type.o gcc `guile-config compile` -c myguile.c -o myguile.o gcc image-type.o myguile.o `guile-config link` -o myguile zwingli:example-smob$ ./myguile guile> make-image #<primitive-procedure make-image> guile> (define i (make-image "Whistler's Mother" 100 100)) guile> i #<image Whistler's Mother> guile> (clear-image i) guile> (clear-image 4) ERROR: In procedure clear-image in expression (clear-image 4): ERROR: Wrong type argument in position 1: 4 ABORT: (wrong-type-arg) Type "(backtrace)" to get more information. guile>
The term "smob" was coined by Aubrey Jaffer, who says it comes from "small object", referring to the fact that only the CDR and part of the CAR of a smob's cell are available for use.
This document was generated on 17 April 1999 using the texi2html translator version 1.51.