3 Memory Allocation

Racket uses both malloc and allocation functions provided by a garbage collector. Embedding/extension C/C++ code may use either allocation method, keeping in mind that pointers to garbage-collectable blocks in malloced memory are invisible (i.e., such pointers will not prevent the block from being garbage-collected).

Racket CGC uses a conservative garbage collector. This garbage collector normally only recognizes pointers to the beginning of allocated objects. Thus, a pointer into the middle of a GC-allocated string will normally not keep the string from being collected. The exception to this rule is that pointers saved on the stack or in registers may point to the middle of a collectable object. Thus, it is safe to loop over an array by incrementing a local pointer variable.

Racket 3m uses a precise garbage collector that moves objects during collection, in which case the C code must be instrumented to expose local pointer bindings to the collector, and to provide tracing procedures for (tagged) records containing pointers. This instrumentation is described further in Cooperating with 3m.

The basic collector allocation functions are:

• scheme_malloc — Allocates collectable memory that may contain pointers to collectable objects; for 3m, the memory must be an array of pointers (though not necessarily to collectable objects). The newly allocated memory is initially zeroed.

• scheme_malloc_atomic — Allocates collectable memory that does not contain pointers to collectable objects. If the memory does contain pointers, they are invisible to the collector and will not prevent an object from being collected. Newly allocated atomic memory is not necessary zeroed.

  Atomic memory is used for strings or other blocks of memory which do not contain pointers. Atomic memory can also be used to store intentionally-hidden pointers.

• scheme_malloc_tagged — Allocates collectable memory that contains a mixture of pointers and atomic data. With the conservative collector, this function is the same as scheme_malloc, but on 3m, the type tag stored at the start of the block is used to determine the size and shape of the object for future garbage collection (as described in Cooperating with 3m).

• scheme_malloc_allow_interior — Allocates an array of pointers such that the array is never moved by the garbage collector and references are allowed into the middle of the block on 3m (and pointers to the middle prevent the block from being collected). Use this procedure sparingly, because small, non-moving objects are handled less efficiently than movable objects by the 3m collector. This procedure is the same as scheme_malloc with the conservative collector, but in the that case, having only a pointer into the interior will not prevent the array from being collected.

• scheme_malloc_atomic_allow_interior — Like scheme_malloc_allow_interior for memory that does not contain pointers.

• scheme_malloc_uncollectable — Allocates uncollectable memory that may contain pointers to collectable objects. There is no way to free the memory. The newly allocated memory is initially zeroed. This function is not available in 3m.

If a Racket extension stores Racket pointers in a global or static variable, then that variable must be registered with scheme_register_extension_global; this makes the pointer visible to the garbage collector. Registered variables need not contain a collectable pointer at all times (even with 3m, but the variable must contain some pointer, possibly uncollectable, at all times). Beware that static or global variables that are not thread-specific (in the OS sense of “thread”) generally do not work with multiple places.

With conservative collection, no registration is needed for the global or static variables of an embedding program, unless it calls scheme_main_setup or scheme_set_stack_base with a non-zero first or second (respectively) argument. In that case, global and static variables containing collectable pointers must be registered with scheme_register_static. The MZ_REGISTER_STATIC macro takes any variable name and registers it with scheme_register_static. The scheme_register_static function can be safely called even when it’s not needed, but it must not be called multiple times for a single memory address.

Collectable memory can be temporarily locked from collection by using the reference-counting function scheme_dont_gc_ptr. On 3m, such locking does not prevent the object from being moved.

Garbage collection can occur during any call into Racket or its allocator, on anytime that Racket has control, except during functions that are documented otherwise. The predicate and accessor macros listed in Standard Types never trigger a collection.

3.1 Cooperating with 3m

To allow 3m’s precise collector to detect and update pointers during garbage collection, all pointer values must be registered with the collector, at least during the times that a collection may occur. The content of a word registered as a pointer must contain either NULL, a pointer to the start of a collectable object, a pointer into an object allocated by scheme_malloc_allow_interior, a pointer to an object currently allocated by another memory manager (and therefore not into a block that is currently managed by the collector), or a pointer to an odd-numbered address (e.g., a Racket fixnum).

Pointers are registered in three different ways:

• Pointers in static variables should be registered with scheme_register_static or MZ_REGISTER_STATIC.

• Pointers in allocated memory are registered automatically when they are in an array allocated with scheme_malloc, etc. When a pointer resides in an object allocated with scheme_malloc_tagged, etc.~the tag at the start of the object identifiers the object’s size and shape. Handling of tags is described in Tagged Objects.

• Local pointers (i.e., pointers on the stack or in registers) must be registered through the MZ_GC_DECL_REG, etc. macros that are described in Local Pointers.

A pointer must never refer to the interior of an allocated object (when a garbage collection is possible), unless the object was allocated with scheme_malloc_allow_interior. For this reason, pointer arithmetic must usually be avoided, unless the variable holding the generated pointer is NULLed before a collection.

IMPORTANT: The SCHEME_SYM_VAL, SCHEME_KEYWORD_VAL, SCHEME_VEC_ELS, and SCHEME_PRIM_CLOSURE_ELS macros produce pointers into the middle of their respective objects, so the results of these macros must not be held during the time that a collection can occur. Incorrectly retaining such a pointer can lead to a crash.

3.1.1 Tagged Objects

As explained in Values and Types, the scheme_make_type function can be used to obtain a new tag for a new type of object. These new types are in relatively short supply for 3m; the maximum tag is 512, and Racket itself uses nearly 300.

After allocating a new tag in 3m (and before creating instances of the tag), a size procedure, a mark procedure, and a fixup procedure must be installed for the tag using GC_register_traversers. A type tag and its associated GC procedures apply to all places, even though specific allocated objects are confined to a particular place.

A size procedure simply takes a pointer to an object with the tag and returns its size in words (not bytes). The gcBYTES_TO_WORDS macro converts a byte count to a word count.

A mark procedure is used to trace references among objects without moving any objects. The procedure takes a pointer to an object, and it should apply the gcMARK macro to every pointer within the object. The mark procedure should return the same result as the size procedure.

A fixup procedure is used to update references to objects after or while they are moved. The procedure takes a pointer to an object, and it should apply the gcFIXUP macro to every pointer within the object; the expansion of this macro takes the address of its argument. The fixup procedure should return the same result as the size procedure.

Depending on the collector’s implementation, the mark or fixup procedure might not be used. For example, the collector may only use the mark procedure and not actually move the object. Or it may use the fixup procedure to mark and move objects at the same time. To dereference an object pointer during a fixup procedure, use GC_fixup_self to convert the address passed to the procedure to refer to the potentially moved object, and use GC_resolve to convert an address that is not yet fixed up to determine the object’s current location.

When allocating a tagged object in 3m, the tag must be installed immediately after the object is allocated—or, at least, before the next possible collection.

3.1.2 Local Pointers

The 3m collector needs to know the address of every local or temporary pointer within a function call at any point when a collection can be triggered. Beware that nested function calls can hide temporary pointers; for example, in

scheme_make_pair(scheme_make_pair(scheme_true, scheme_false),

scheme_make_pair(scheme_false, scheme_true))

the result from one scheme_make_pair call is on the stack or in a register during the other call to scheme_make_pair; this pointer must be exposed to the garbage collection and made subject to update. Simply changing the code to

tmp = scheme_make_pair(scheme_true, scheme_false);

scheme_make_pair(tmp,

scheme_make_pair(scheme_false, scheme_true))

does not expose all pointers, since tmp must be evaluated before the second call to scheme_make_pair. In general, the above code must be converted to the form

tmp1 = scheme_make_pair(scheme_true, scheme_false);

tmp2 = scheme_make_pair(scheme_true, scheme_false);

scheme_make_pair(tmp1, tmp2);

and this is converted form must be instrumented to register tmp1 and tmp2. The final result might be

{

Scheme_Object *tmp1 = NULL, *tmp2 = NULL, *result;

MZ_GC_DECL_REG(2);

MZ_GC_VAR_IN_REG(0, tmp1);

MZ_GC_VAR_IN_REG(1, tmp2);

MZ_GC_REG();

tmp1 = scheme_make_pair(scheme_true, scheme_false);

tmp2 = scheme_make_pair(scheme_true, scheme_false);

result = scheme_make_pair(tmp1, tmp2);

MZ_GC_UNREG();

return result;

}

Notice that result is not registered above. The MZ_GC_UNREG macro cannot trigger a garbage collection, so the result variable is never live during a potential collection. Note also that tmp1 and tmp2 are initialized with NULL, so that they always contain a pointer whenever a collection is possible.

The MZ_GC_DECL_REG macro expands to a local-variable declaration to hold information for the garbage collector. The argument is the number of slots to provide for registration. Registering a simple pointer requires a single slot, whereas registering an array of pointers requires three slots. For example, to register a pointer tmp and an array of 10 char*s:

{

Scheme_Object *tmp1 = NULL;

char *a[10];

int i;

MZ_GC_DECL_REG(4);

MZ_GC_ARRAY_VAR_IN_REG(0, a, 10);

MZ_GC_VAR_IN_REG(3, tmp1);

/* Clear a before a potential GC: */

for (i = 0; i < 10; i++) a[i] = NULL;

...

f(a);

...

}

The MZ_GC_ARRAY_VAR_IN_REG macro registers a local array given a starting slot, the array variable, and an array size. The MZ_GC_VAR_IN_REG macro takes a slot and simple pointer variable. A local variable or array must not be registered multiple times.

In the above example, the first argument to MZ_GC_VAR_IN_REG is 3 because the information for a uses the first three slots. Even if a is not used after the call to f, a must be registered with the collector during the entire call to f, because f presumably uses a until it returns.

The name used for a variable need not be immediate. Structure members can be supplied as well:

{

struct { void *s; int v; void *t; } x = {NULL, 0, NULL};

MZ_GC_DECL_REG(2);

MZ_GC_VAR_IN_REG(0, x.s);

MZ_GC_VAR_IN_REG(0, x.t);

...

}

In general, the only constraint on the second argument to MZ_GC_VAR_IN_REG or MZ_GC_ARRAY_VAR_IN_REG is that & must produce the relevant address, and that address must be on the stack.

Pointer information is not actually registered with the collector until the MZ_GC_REG macro is used. The MZ_GC_UNREG macro de-registers the information. Each call to MZ_GC_REG must be balanced by one call to MZ_GC_UNREG.

Pointer information need not be initialized with MZ_GC_VAR_IN_REG and MZ_GC_ARRAY_VAR_IN_REG before calling MZ_GC_REG, and the set of registered pointers can change at any time—as long as all relevant pointers are registered when a collection might occur. The following example recycles slots and completely de-registers information when no pointers are relevant. The example also illustrates how MZ_GC_UNREG is not needed when control escapes from the function, such as when scheme_signal_error escapes.

{

Scheme_Object *tmp1 = NULL, *tmp2 = NULL;

mzchar *a, *b;

MZ_GC_DECL_REG(2);

MZ_GC_VAR_IN_REG(0, tmp1);

MZ_GC_VAR_IN_REG(1, tmp2);

tmp1 = scheme_make_utf8_string("foo");

MZ_GC_REG();

tmp2 = scheme_make_utf8_string("bar");

tmp1 = scheme_append_char_string(tmp1, tmp2);

if (SCHEME_FALSEP(tmp1))

scheme_signal_error("shouldn't happen!");

a = SCHEME_CHAR_VAL(tmp1);

MZ_GC_VAR_IN_REG(0, a);

tmp2 = scheme_make_pair(scheme_read_bignum(a, 0, 10), tmp2);

MZ_GC_UNREG();

if (SCHEME_INTP(tmp2)) {

return 0;

}

MZ_GC_REG();

tmp1 = scheme_make_pair(scheme_read_bignum(a, 0, 8), tmp2);

MZ_GC_UNREG();

return tmp1;

}

A MZ_GC_DECL_REG can be used in a nested block to hold declarations for the block’s variables. In that case, the nested MZ_GC_DECL_REG must have its own MZ_GC_REG and MZ_GC_UNREG calls.

{

Scheme_Object *accum = NULL;

MZ_GC_DECL_REG(1);

MZ_GC_VAR_IN_REG(0, accum);

MZ_GC_REG();

accum = scheme_make_pair(scheme_true, scheme_null);

{

Scheme_Object *tmp = NULL;

MZ_GC_DECL_REG(1);

MZ_GC_VAR_IN_REG(0, tmp);

MZ_GC_REG();

tmp = scheme_make_pair(scheme_true, scheme_false);

accum = scheme_make_pair(tmp, accum);

MZ_GC_UNREG();

}

accum = scheme_make_pair(scheme_true, accum);

MZ_GC_UNREG();

return accum;

}

Variables declared in a local block can also be registered together with variables from an enclosing block, but the local-block variable must be unregistered before it goes out of scope. The MZ_GC_NO_VAR_IN_REG macro can be used to unregister a variable or to initialize a slot as having no variable.

{

Scheme_Object *accum = NULL;

MZ_GC_DECL_REG(2);

MZ_GC_VAR_IN_REG(0, accum);

MZ_GC_NO_VAR_IN_REG(1);

MZ_GC_REG();

accum = scheme_make_pair(scheme_true, scheme_null);

{

Scheme_Object *tmp = NULL;

MZ_GC_VAR_IN_REG(1, tmp);

tmp = scheme_make_pair(scheme_true, scheme_false);

accum = scheme_make_pair(tmp, accum);

MZ_GC_NO_VAR_IN_REG(1);

}

accum = scheme_make_pair(scheme_true, accum);

MZ_GC_UNREG();

return accum;

}

The MZ_GC_ macros all expand to nothing when MZ_PRECISE_GC is not defined, so the macros can be placed into code to be compiled for both conservative and precise collection.

The MZ_GC_REG and MZ_GC_UNREG macros must never be used in an OS thread other than Racket’s thread.

3.1.3 Local Pointers and raco ctool --xform

When raco ctool is run with the --xform flag and a source C program, it produces a C program that is instrumented in the way described in the previous section (but with a slightly different set of macros). For each input file "name.c", the transformed output is "name.3m.c".

The --xform mode for raco ctool does not change allocation calls, nor does it generate size, mark, or fixup procedures. It merely converts the code to register local pointers.

Furthermore, the --xform mode for raco ctool does not handle all of C. It’s ability to rearrange compound expressions is particularly limited, because --xform merely converts expression text heuristically instead of parsing C. A future version of the tool will correct such problems. For now, raco ctool in --xform mode attempts to provide reasonable error messages when it is unable to convert a program, but beware that it can miss cases. To an even more limited degree, --xform can work on C++ code. Inspect the output of --xform mode to ensure that your code is correctly instrumented.

Some specific limitations:

• The body of a for, while, or do loop must be surrounded with curly braces. (A conversion error is normally reported, otherwise.)

• Function calls may not appear on the right-hand side of an assignment within a declaration block. (A conversion error is normally reported if such an assignment is discovered.)

• Multiple function calls in ... ? ... : ... cannot be lifted. (A conversion error is normally reported, otherwise.)

• In an assignment, the left-hand side must be a local or static variable, not a field selection, pointer dereference, etc. (A conversion error is normally reported, otherwise.)

• The conversion assumes that all function calls use an immediate name for a function, as opposed to a compound expression as in s->f(). The function name need not be a top-level function name, but it must be bound either as an argument or local variable with the form type id; the syntax ret_type (*id)(...) is not recognized, so bind the function type to a simple name with typedef, first: typedef ret_type (*type)(...); .... type id.

• Arrays and structs must be passed by address, only.

• GC-triggering code must not appear in system headers.

• Pointer-comparison expressions are not handled correctly when either of the compared expressions includes a function call. For example, a() == b() is not converted correctly when a and b produce pointer values.

• Passing the address of a local pointer to a function works only when the pointer variable remains live after the function call.

• A return; form can get converted to { stmt; return; };, which can break an if (...) return; else ... pattern.

• Local instances of union types are generally not supported.

• Pointer arithmetic cannot be converted away, and is instead reported as an error.

3.1.4 Guiding raco ctool --xform

The following macros can be used (with care!) to navigate --xform around code that it cannot handle:

• XFORM_START_SKIP and XFORM_END_SKIP: code between these two statements is ignored by the transform tool, except to tokenize it.

  Example:

  int foo(int c, ...) {

  int r = 0;

  XFORM_START_SKIP;

  {

  /* va plays strange tricks that confuse xform */

  va_list args;

  va_start(args, c);

  while (c--) {

  r += va_arg(args, int);

  }

  }

  XFORM_END_SKIP;

  return r;

  }

  These macros can also be used at the top level, outside of any function. Since they have to be terminated by a semi-colon, however, top-level uses usually must be wrapped with #ifdef MZ_PRECISE_GC and #endif; a semi-colon by itself at the top level is not legal in C.

• XFORM_SKIP_PROC: annotate a function so that its body is skipped in the same way as bracketing it with XFORM_START_SKIP and XFORM_END_SKIP.

  Example:

  int foo(int c, ...) XFORM_END_SKIP {

  }

• XFORM_HIDE_EXPR: a macro that takes wraps an expression to disable processing of the expression.

  Example:

  int foo(int c, ...) {

  int r = 0;

  {

  /* va plays strange tricks that confuse xform */

  XFORM_CAN_IGNORE va_list args; /* See below */

  XFORM_HIDE_EXPR(va_start(args, c));

  while (c--) {

  r += XFORM_HIDE_EXPR(va_arg(args, int));

  }

  }

  return r;

  }

• XFORM_CAN_IGNORE: a macro that acts like a type modifier (must appear first) to indicate that a declared variable can be treated as atomic. See above for an example.

• XFORM_START_SUSPEND and XFORM_END_SUSPEND: for use at the top level (outside of any function definition), and similar to XFORM_START_SKIP and XFORM_END_SKIP in that function and class bodies are not transformed. Type and prototype information is still collected for use by later transformations, however. These forms must be terminated by a semi-colon.

• XFORM_START_TRUST_ARITH and XFORM_END_TRUST_ARITH: for use at the top level (outside of any function definition) to disable warnings about pointer arithmetic. Use only when you’re absolutely certain that the garbage collector cannot be pointers offset into the middle of a collectable object. These forms must be terminated by a semi-colon.

• XFORM_TRUST_PLUS: a replacement for + that does not trigger pointer-arithmetic warnings. Use with care.

• XFORM_TRUST_MINUS: a replacement for - that does not trigger pointer-arithmetic warnings. Use with care.

3.2 Memory Functions

In addition to global variables, this function can be used to register any permanent memory that the collector would otherwise treat as atomic. A garbage collection can occur during the registration.

The Scheme_Env_Main type is defined as follows:

typedef int (*Scheme_Env_Main)(Scheme_Env *env,

int argc, char **argv);

The result of main is the result of scheme_main_setup.

If no_auto_statics is non-zero, then static variables must be explicitly registered with the garbage collector; see Memory Allocation for more information.

The data argument is passed through to main, where the Scheme_Nested_Main type is defined as follows:

typedef int (*Scheme_Nested_Main)(void *data);

The following example shows a typical use for setting the stack base for CGC:

int main(int argc, char **argv) {

int dummy;

scheme_set_stack_base(&dummy, 0);

real_main(argc, argv); /* calls scheme_basic_env(), etc. */

}

On 3m, the above code does not quite work, because stack_addr must be the beginning or end of a local-frame registration. Worse, in CGC or 3m, if real_main is declared static, the compiler may inline it and place variables containing collectable values deeper in the stack than dummy. To avoid these problems, use scheme_main_setup or scheme_main_stack_setup, instead.

If stack_end is NULL, then the stack end is computed automatically: the stack size assumed to be the limit reported by getrlimit on Unix and Mac OS X, or it is assumed to be 1 MB on Windows; if this size is greater than 8 MB, then 8 MB is assumed, instead; the size is decremented by 50000 bytes to cover a large margin of error; finally, the size is subtracted from (for stacks that grow down) or added to (for stacks that grow up) the stack base in stack_addr or the automatically computed stack base. Note that the 50000-byte margin of error is assumed to cover the difference between the actual stack start and the reported stack base, in addition to the margin needed for detecting and handling stack overflow.

The tls_index argument must be 0. It is currently ignored, but a future version may use the argument to allow declaration of the thread-local variable in a dynamically linked DLL.

The macro MZ_REGISTER_STATIC can be used directly on a static variable. It expands to a comment if statics need not be registered, and a call to scheme_register_static (with the address of the static variable) otherwise.

This function is not available in 3m.

This function is not available in 3m.

The fnl_proc type is not actually defined, but it is equivalent to

typedef void (*fnl_proc)(void *p, void *data)

The f argument is the callback function; when it is called, it will be passed the value p and the data pointer data; data can be anything — it is only passed on to the callback function. If oldf and olddata are not NULL, then *oldf and *olddata are filled with the old callback information (f and data will override this old callback).

To remove a registered finalizer, pass NULL for f and data.

Note: registering a callback not only keeps p from collection until the callback is invoked, but it also keeps data reachable until the callback is invoked.

See scheme_register_finalizer, above, for information about the arguments.

To remove an added finalizer, use scheme_subtract_finalizer.

See scheme_register_finalizer, above, for information about the arguments.

There is currently no facility to remove a will-like finalizer.

This function keeps a reference count on the pointers it registers, so two calls to scheme_dont_gc_ptr for the same p should be balanced with two calls to scheme_gc_ptr_ok.

When the PLTDISABLEGC environment variable is set, then racket initializes the internal counter to 1 to initially disable garbage collection.

Each of the three procedures takes a pointer and returns an integer:

typedef int (*Size_Proc)(void *obj);

typedef int (*Mark_Proc)(void *obj);

typedef int (*Fixup_Proc)(void *obj);

If the result of the size procedure is a constant, then pass a non-zero value for is_const_size. If the mark and fixup procedures are no-ops, then pass a non-zero value for is_atomic.

A description is a vector of vectors, where each of the inner vectors describes a single call, and the calls are performed in sequence. Each call vector starts with a symbol that indicates the protocol of the foreign function to be called. The following protocols are supported:

• 'ptr_ptr_ptr->void corresponds to void (*)(void*, void*, void*).

• 'ptr_ptr_ptr_int->void corresponds to void (*)(void*, void*, void*, int).

• 'ptr_ptr_float->void corresponds to void (*)(void*, void*, float).

• 'ptr_ptr_double->void corresponds to void (*)(void*, void*, double).

• 'ptr_ptr_ptr_int_int_int_int_int_int_int_int_int->void corresponds to void (*)(void*, void*, void*, int, int, int, int, int, int, int, int, int).

• 'osapi_ptr_int->void corresponds to void (*)(void*, int), but using the stdcall calling convention on Windows.

• 'osapi_ptr_ptr->void corresponds to void (*)(void*, void*), but using the stdcall calling convention on Windows.

• 'osapi_ptr_int_int_int_int_ptr_int_int_long->void corresponds to void (*)(void*, int, int, int, int, void*, int, int, long), but using the stdcall calling convention on Windows.

After the protocol symbol, the vector should contain a pointer to a foreign function and then an element for each of the function’s arguments. Pointer values are represented as for the _pointer representation defined by ffi/unsafe.

The result is a key for use with scheme_remove_gc_callback. If the key becomes inaccessible, then the callback will be removed automatically (but beware that the pre-callback will have executed and the post-callback will not have executed).