The following consists of step-by-step instructions for building and using the collector. We'll assume a Linux/gcc platform and a single-threaded application. The green text contains information about other platforms or scenarios. It can be skipped, especially on first reading.
If you have not so yet, unpack the collector and enter the newly created directory with:
tar xvfz gc-<version>.tar.gz
cd gc-<version>
You can configure, build, and install the collector in a private directory, say /home/xyz/gc, with the following commands:
./configure --prefix=/home/xyz/gc --disable-threads
make
make check
make install
Here the make check
command is optional, but highly recommended. It runs
a basic correctness test which usually takes well under a minute.
On non-Unix, non-Linux platforms, the collector is usually built by copying
the appropriate makefile (see the platform-specific README in docs/platforms
folder in the distribution) to the file Makefile
, and then typing make
(or
nmake
or ...). This builds the library in the source tree. You may want
to move it and the files in the include directory to a more convenient place.
If you use a makefile that does not require running a configure script, you should first look at the makefile, and adjust any options that are documented there.
If your platform provides a make
utility, that is generally preferred
to platform- and compiler- dependent "project" files. (At least that is the
strong preference of the would-be maintainer of those project files.)
If you do not need thread support, configure the collector with:
--disable-threads
Alternatively, if your target is a real old-fashioned uniprocessor (no
"hyperthreading", etc.), you may just want to turn off parallel marking with
--disable-parallel-mark
.
You will need to include the C++ support, which unfortunately tends to be
among the least portable parts of the collector, since it seems to rely
on some corner cases of the language. On Linux, it suffices to add
--enable-cplusplus
to the configure options.
You will need to include "gc.h" at the beginning of every file that allocates
memory through the garbage collector. Call GC_MALLOC
wherever you would have
call malloc
. This initializes memory to zero like calloc
; there is no need
to explicitly clear the result.
If you know that an object will not contain pointers to the garbage-collected
heap, and you don't need it to be initialized, call GC_MALLOC_ATOMIC
instead.
A function GC_FREE
is provided but need not be called. For very small
objects, your program will probably perform better if you do not call it, and
let the collector do its job.
A GC_REALLOC
function behaves like the C library realloc
. It allocates
uninitialized pointer-free memory if the original object was allocated that
way.
The following program loop.c
is a trivial example:
#include "gc.h"
#include <assert.h>
#include <stdio.h>
int main(void) {
int i;
GC_INIT();
for (i = 0; i < 10000000; ++i) {
int **p = (int **) GC_MALLOC(sizeof(int *));
int *q = (int *) GC_MALLOC_ATOMIC(sizeof(int));
assert(*p == 0);
*p = (int *) GC_REALLOC(q, 2 * sizeof(int));
if (i % 100000 == 0)
printf("Heap size = %lu bytes\n",
(unsigned long)GC_get_heap_size());
}
return 0;
}
It is usually best not to mix garbage-collected allocation with the system
malloc
-free
. If you do, you need to be careful not to store pointers
to the garbage-collected heap in memory allocated with the system malloc
.
On some other platforms it is necessary to call GC_INIT
from the main
program, which is presumed to be part of the main executable, not a dynamic
library. This can never hurt, and is thus generally good practice.
For a multi-threaded program, some more rules apply:
- Files that either allocate through the GC or make thread-related calls
should first define the macro
GC_THREADS
, and then includegc.h
. On some platforms this will redefine some threads primitives, e.g. to let the collector keep track of thread creation.
In the case of C++, you need to be especially careful not to store pointers to the garbage-collected heap in areas that are not traced by the collector. The collector includes some alternate interfaces to make that easier.
Additional debug checks can be performed by defining GC_DEBUG
before
including gc.h
. Additional options are available if the collector is also
built with --enable-gc-debug
and all allocations are performed with
GC_DEBUG
defined.
You may be able to build the collector with --enable-redirect-malloc
and set
the LD_PRELOAD
environment variable to point to the resulting library, thus
replacing the standard malloc
with its garbage-collected counterpart. This
is rather platform dependent. See the GC leak detection documentation for
some more details.
The above application loop.c
test program can be compiled and linked with:
cc -I/home/xyz/gc/include loop.c /home/xyz/gc/lib/libgc.a -o loop
The -I
option directs the compiler to the right include directory. In this
case, we list the static library directly on the compile line; the dynamic
library could have been used instead, provided we arranged for the dynamic
loader to find it, e.g. by setting LD_LIBRARY_PATH
.
On pthread platforms, you will of course also have to link with -lpthread
,
and compile with any thread-safety options required by your compiler. On some
platforms, you may also need to link with -ldl
or -lrt
. Looking
at tools/threadlibs.c
should give you the appropriate list if a plain
-lpthread
does not work.
The executable can of course be run normally, e.g. by typing:
./loop
The operation of the collector is affected by a number of environment
variables. For example, setting GC_PRINT_STATS
produces some GC statistics
on stdout. See README.environment file for the details.