1.1 Motivation for writing libtool

Since early 1995, several different GNU developers have recognized the importance of having shared library support for their packages. The primary motivation for such a change is to encourage modularity and reuse of code (both conceptually and physically) in GNU programs.

Such a demand means that the way libraries are built in GNU packages needs to be general, to allow for any library type the package installer might want. The problem is compounded by the absence of a standard procedure for creating shared libraries on different platforms.

The following sections outline the major issues facing shared library support in GNU, and how I propose that shared library support could be standardized with libtool.

The following specifications were used in developing and evaluating this system:

1. The system must be as elegant as possible.
2. The system must be fully integrated with the GNU Autoconf and Automake utilities, so that it will be easy for GNU maintainers to use. However, the system must not require these tools, so that it can be used by non-GNU packages.
3. Portability to other (non-GNU) architectures and tools is desirable.

1.2 Implementation issues

The following issues need to be addressed in any reusable shared library system, specifically libtool:

1. The package installer should be able to control what sort of libraries are built.
2. It can be tricky to run dynamically linked programs whose libraries have not yet been installed. LD_LIBRARY_PATH must be set properly (if it is supported), or programs fail to run.
3. The system must operate consistently even on hosts which don’t support shared libraries.
4. The commands required to build shared libraries may differ wildly from host to host. These need to be determined at configure time in a consistent way.
5. It is not always obvious with which suffix a shared library should be installed. This makes it difficult for Makefile rules, since they generally assume that file names are the same from host to host.
6. The system needs a simple library version number abstraction, so that shared libraries can be upgraded in place. The programmer should be informed how to design the interfaces to the library to maximize binary compatibility.
7. The install Makefile target should warn the package installer to set the proper environment variables (LD_LIBRARY_PATH or equivalent), or run ldconfig(8).

1.3 Other implementations

I have investigated several different implementations of systems that build shared libraries as part of a free software package. At first, I made notes on the features of each of these packages for comparison purposes.

Now it is clear that none of these packages have documented the details of shared library systems that libtool requires. So, other packages have been more or less abandoned as influences.

1.4 A postmortem analysis of other implementations

In all fairness, each of the implementations that I examined do the job that they were intended to do, for a number of different host systems. However, none of these solutions seem to function well as a generalized, reusable component.

Most were too complex for me to use (much less modify) without understanding exactly what the implementation does, and they were generally not documented.

I think the main problem is that different vendors have different views of what libraries are, and none of the packages I examined seemed to be confident enough to settle on a single paradigm that just works.

Ideally, libtool would be a standard that would be implemented as series of extensions and modifications to existing library systems to make them work consistently. However, I don’t have the time or power to convince operating system developers to mend their evil ways, and I want to build shared libraries right now, even on buggy, broken, confused operating systems.

For this reason, I have designed libtool as an independent shell script. It isolates the problems and inconsistencies in library building that plague Makefile writers by wrapping the compiler suite on different platforms with a consistent, powerful interface.

I hope that libtool will be useful to and used by the GNU community, and that the lessons I’ve learned in writing it will be taken up and implemented by designers of library systems.

3.1 Creating object files

To create an object file from a source file, the compiler is invoked with the ‘-c’ flag (and any other desired flags):

burger$ gcc -g -O -c main.c
burger$

The above compiler command produces an object file, main.o, from the source file main.c.

For most library systems, creating object files that become part of a static library is as simple as creating object files that are linked to form an executable:

burger$ gcc -g -O -c foo.c
burger$ gcc -g -O -c hello.c
burger$

Shared libraries, however, may only be built from position-independent code (PIC). So, special flags must be passed to the compiler to tell it to generate PIC rather than the standard position-dependent code.

Since this is a library implementation detail, libtool hides the complexity of PIC compiler flags by using separate library object files (which end in ‘.lo’ instead of ‘.o’). On systems without shared libraries (or without special PIC compiler flags), these library object files are identical to “standard” object files.

To create library object files for foo.c and hello.c, simply invoke libtool with the standard compilation command as arguments (see Compile mode):

a23$ libtool gcc -g -O -c foo.c
gcc -g -O -c foo.c
ln -s foo.o foo.lo
a23$ libtool gcc -g -O -c hello.c
gcc -g -O -c hello.c
ln -s hello.o hello.lo
a23$

Note that libtool creates two object files for each invocation. The ‘.lo’ file is a library object, and the ‘.o’ file is a standard object file. On ‘a23’, these files are identical, because only static libraries are supported.

On shared library systems, libtool automatically inserts the PIC generation flags into the compilation command, so that the library object and the standard object differ:

burger$ libtool gcc -g -O -c foo.c
gcc -g -O -c -fPIC -DPIC foo.c
mv -f foo.o foo.lo
gcc -g -O -c foo.c
burger$ libtool gcc -g -O -c hello.c
gcc -g -O -c -fPIC -DPIC hello.c
mv -f hello.o hello.lo
gcc -g -O -c hello.c
burger$

3.2 Linking libraries

Without libtool, the programmer would invoke the ar command to create a static library:

burger$ ar cru libhello.a hello.o foo.o
burger$

But of course, that would be too simple, so many systems require that you run the ranlib command on the resulting library (to give it better karma, or something):

burger$ ranlib libhello.a
burger$

It seems more natural to use the C compiler for this task, given libtool’s “libraries are programs” approach. So, on platforms without shared libraries, libtool simply acts as a wrapper for the system ar (and possibly ranlib) commands.

Again, the libtool library name differs from the standard name (it has a ‘.la’ suffix instead of a ‘.a’ suffix). The arguments to libtool are the same ones you would use to produce an executable named libhello.la with your compiler (see Link mode):

burger$ libtool gcc -g -O -o libhello.la foo.o hello.o
libtool: cannot build libtool library `libhello.la' from non-libtool \
                objects
burger$

Aha! Libtool caught a common error… trying to build a library from standard objects instead of library objects. This doesn’t matter for static libraries, but on shared library systems, it is of great importance.

So, let’s try again, this time with the library object files:¹

a23$ libtool gcc -g -O -o libhello.la foo.lo hello.lo -lm
libtool: you must specify an installation directory with `-rpath'
a23$

Argh. Another complication in building shared libraries is that we need to specify the path to the directory in which they (eventually) will be installed. So, we try again, with an rpath setting of /usr/local/lib:

a23$ libtool gcc -g -O -o libhello.la foo.lo hello.lo \
                -rpath /usr/local/lib -lm
mkdir .libs
ar cru .libs/libhello.a foo.o hello.o
ranlib .libs/libhello.a
creating libhello.la
a23$

Now, let’s try the same trick on the shared library platform:

burger$ libtool gcc -g -O -o libhello.la foo.lo hello.lo \
                -rpath /usr/local/lib -lm
mkdir .libs
ld -Bshareable -o .libs/libhello.so.0.0 foo.lo hello.lo -lm
ar cru .libs/libhello.a foo.o hello.o
ranlib .libs/libhello.a
creating libhello.la
burger$

Now that’s significantly cooler… libtool just ran an obscure ld command to create a shared library, as well as the static library.

Note how libtool creates extra files in the .libs subdirectory, rather than the current directory. This feature is to make it easier to clean up the build directory, and to help ensure that other programs fail horribly if you accidentally forget to use libtool when you should.

3.3 Linking executables

If you choose at this point to install the library (put it in a permanent location) before linking executables against it, then you don’t need to use libtool to do the linking. Simply use the appropriate ‘-L’ and ‘-l’ flags to specify the library’s location.

Some system linkers insist on encoding the full directory name of each shared library in the resulting executable. Libtool has to work around this misfeature by special magic to ensure that only permanent directory names are put into installed executables.

The importance of this bug must not be overlooked: it won’t cause programs to crash in obvious ways. It creates a security hole, and possibly even worse, if you are modifying the library source code after you have installed the package, you will change the behaviour of the installed programs!

So, if you want to link programs against the library before you install it, you must use libtool to do the linking.

Here’s the old way of linking against an uninstalled library:

burger$ gcc -g -O -o hell.old main.o libhello.a -lm
burger$

Libtool’s way is almost the same² (see Link mode):

a23$ libtool gcc -g -O -o hell main.o libhello.la -lm
gcc -g -O -o hell main.o ./.libs/libhello.a -lm
a23$

That looks too simple to be true. All libtool did was transform libhello.la to ./.libs/libhello.a, but remember that ‘a23’ has no shared libraries.

On ‘burger’ the situation is different:

burger$ libtool gcc -g -O -o hell main.o libhello.la -lm
gcc -g -O -o .libs/hell main.o -L./.libs -R/usr/local/lib -lhello -lm
creating hell
burger$

Notice that the executable, hell, was actually created in the .libs subdirectory. Then, a wrapper script was created in the current directory.

On NetBSD 1.2, libtool encodes the installation directory of libhello, by using the ‘-R/usr/local/lib’ compiler flag. Then, the wrapper script guarantees that the executable finds the correct shared library (the one in ./.libs) until it is properly installed.

Let’s compare the two different programs:

burger$ time ./hell.old
Welcome to GNU Hell!
** This is not GNU Hello.  There is no built-in mail reader. **
        0.21 real         0.02 user         0.08 sys
burger$ time ./hell
Welcome to GNU Hell!
** This is not GNU Hello.  There is no built-in mail reader. **
        0.63 real         0.09 user         0.59 sys
burger$

The wrapper script takes significantly longer to execute, but at least the results are correct, even though the shared library hasn’t been installed yet.

So, what about all the space savings that shared libraries are supposed to yield?

burger$ ls -l hell.old libhello.a
-rwxr-xr-x  1 gord  gord  15481 Nov 14 12:11 hell.old
-rw-r--r--  1 gord  gord   4274 Nov 13 18:02 libhello.a
burger$ ls -l .libs/hell .libs/libhello.*
-rwxr-xr-x  1 gord  gord  11647 Nov 14 12:10 .libs/hell
-rw-r--r--  1 gord  gord   4274 Nov 13 18:44 .libs/libhello.a
-rwxr-xr-x  1 gord  gord  12205 Nov 13 18:44 .libs/libhello.so.0.0
burger$

Well, that sucks. Maybe I should just scrap this project and take up basket weaving.

Actually, it just proves an important point: shared libraries incur overhead because of their (relative) complexity. In this situation, the price of being dynamic is eight kilobytes, and the payoff is about four kilobytes. So, having a shared libhello won’t be an advantage until we link it against at least a few more programs.

3.4 Debugging executables

If hell was a complicated program, you would certainly want to test and debug it before installing it on your system. In the above section, you saw how it the libtool wrapper script makes it possible to run the program directly, but unfortunately, it interferes with the debugger:

burger$ gdb hell
GDB is free software and you are welcome to distribute copies of it
 under certain conditions; type "show copying" to see the conditions.
There is absolutely no warranty for GDB; type "show warranty" for details.
GDB 4.16 (i386-unknown-netbsd), Copyright 1996 Free Software Foundation, Inc...

"hell": not in executable format: File format not recognized

(gdb) quit
burger$

Sad. It doesn’t work because GDB isn’t doesn’t know where the executable lives. So, let’s try again, by invoking GDB directly on the executable:

burger$ gdb .libs/hell
trick:/home/src/libtool/demo$ gdb .libs/hell
GDB is free software and you are welcome to distribute copies of it
 under certain conditions; type "show copying" to see the conditions.
There is absolutely no warranty for GDB; type "show warranty" for details.
GDB 4.16 (i386-unknown-netbsd), Copyright 1996 Free Software Foundation, Inc...
(gdb) break main
Breakpoint 1 at 0x8048547: file main.c, line 29.
(gdb) run
Starting program: /home/src/libtool/demo/.libs/hell
/home/src/libtool/demo/.libs/hell: can't load library 'libhello.so.2'

Program exited with code 020.
(gdb) quit
burger$

Argh. Now GDB complains because it cannot find the shared library that hell is linked against. So, we must use libtool in order to properly set the library path and run the debugger. Fortunately, we can forget all about the .libs directory, and just run it on the executable wrapper (see Execute mode):

burger$ libtool gdb hell
GDB is free software and you are welcome to distribute copies of it
 under certain conditions; type "show copying" to see the conditions.
There is absolutely no warranty for GDB; type "show warranty" for details.
GDB 4.16 (i386-unknown-netbsd), Copyright 1996 Free Software Foundation, Inc...
(gdb) break main
Breakpoint 1 at 0x8048547: file main.c, line 29.
(gdb) run
Starting program: /home/src/libtool/demo/.libs/hell

Breakpoint 1, main (argc=1, argv=0xbffffc40) at main.c:29
29	  printf ("Welcome to GNU Hell!\n");
(gdb) quit
The program is running.  Quit anyway (and kill it)? (y or n) y
burger$

3.5 Installing libraries

Installing libraries on a non-libtool system is quite straightforward… just copy them into place:³

burger$ su
Password: ********
burger# cp libhello.a /usr/local/lib/libhello.a
burger#

Oops, don’t forget the ranlib command:

burger# ranlib /usr/local/lib/libhello.a
burger#

Libtool installation is quite simple, as well. Just use the install or cp command that you normally would (see Install mode):

a23# libtool cp libhello.la /usr/local/lib/libhello.la
cp libhello.la /usr/local/lib/libhello.la
cp .libs/libhello.a /usr/local/lib/libhello.a
ranlib /usr/local/lib/libhello.a
a23#

Note that the libtool library libhello.la is also installed, for informational purposes, and to help libtool with uninstallation (see Uninstall mode).

Here is the shared library example:

burger# libtool install -c libhello.la /usr/local/lib/libhello.la
install -c .libs/libhello.so.0.0 /usr/local/lib/libhello.so.0.0
install -c libhello.la /usr/local/lib/libhello.la
install -c .libs/libhello.a /usr/local/lib/libhello.a
ranlib /usr/local/lib/libhello.a
burger#

It is safe to specify the ‘-s’ (strip symbols) flag if you use a BSD-compatible install program when installing libraries. Libtool will either ignore the ‘-s’ flag, or will run a program that will strip only debugging and compiler symbols from the library.

Once the libraries have been put in place, there may be some additional configuration that you need to do before using them. First, you must make sure that where the library is installed actually agrees with the ‘-rpath’ flag you used to build it.

Then, running ‘libtool -n --finish libdir’ can give you further hints on what to do (see Finish mode):

burger# libtool -n --finish /usr/local/lib
ldconfig -m /usr/local/lib
To link against installed libraries in LIBDIR, users may have to:
   - add LIBDIR to their `LD_LIBRARY_PATH' environment variable
   - use the `-LLIBDIR' linker flag
burger#

After you have completed these steps, you can go on to begin using the installed libraries. You may also install any executables that depend on libraries you created.

3.6 Installing executables

If you used libtool to link any executables against uninstalled libtool libraries (see Linking executables), you need to use libtool to install the executables after the libraries have been installed (see Installing libraries).

So, for our Ultrix example, we would run:

a23# libtool install -c hell /usr/local/bin/hell
install -c hell /usr/local/bin/hell
a23#

On shared library systems, libtool just ignores the wrapper script and installs the correct binary:

burger# libtool install -c hell /usr/local/bin/hell
install -c .libs/hell /usr/local/bin/hell
burger#

3.7 Linking static libraries

Why return to ar and ranlib silliness when you’ve had a taste of libtool? Well, sometimes it is desirable to create a static archive that can never be shared. The most frequent case is when you have a “convenience library” that is a collection of related object files without a really nice interface.

To do this, you should ignore libtool entirely, and just use the old ar and ranlib commands to create a static library.

If you want to install the library (but you probably don’t), then you may use libtool if you want:

burger$ libtool ./install-sh -c libhello.a /local/lib/libhello.a
./install-sh -c libhello.a /local/lib/libhello.a
ranlib /local/lib/libhello.a
burger$

Using libtool for static library installation protects your library from being accidentally stripped (if the installer used the ‘-s’ flag), as well as automatically running the correct ranlib command.

Another common situation where static linking is desirable is in creating a standalone binary. Use libtool to do the linking and add the ‘-all-static’ flag.

4.1 Compile mode

For ‘compile’ mode, mode-args is a compiler command to be used in creating a ‘standard’ object file. These arguments should begin with the name of the C compiler, and contain the ‘-c’ compiler flag so that only an object file is created.

Libtool determines the name of the output file by removing the directory component from the source file name, then substituting the C source code suffix ‘.c’ with the library object suffix, ‘.lo’.

If shared libraries are being built, any necessary PIC generation flags are substituted into the compilation command.

Note that the ‘-o’ option is not supported for compile mode, because it cannot be implemented properly for all platforms. It is far easier just to change your Makefiles to create all the output files in the current working directory.

4.2 Link mode

‘link’ mode links together object files (including library objects) to form another library or to create an executable program.

mode-args consist of a command using the C compiler to create an output file (with the ‘-o’ flag) from several object files.

The following components of mode-args are treated specially:

‘-all-static’

If output-file is a program, then do not link it against any shared libraries at all. If output-file is a library, then only create a static library.

‘-dlopen file’

Same as ‘-dlpreopen file’, if native dlopening is not supported on the host platform (see Dlopened modules). Otherwise, no effect.

‘-dlpreopen file’

Link file into the output program, and add its symbols to dld_preloaded_symbols (see Dlpreopening).

‘-export-dynamic’

Allow symbols from output-file to be resolved with dlsym(3) (see Dlopened modules).

‘-Llibdir’

Search libdir for required libraries that have already been installed.

‘-lname’

output-file requires the installed library libname. This option is required even when output-file is not an executable.

‘-no-undefined’

Declare that output-file does not depend on any other libraries. Some platforms cannot create shared libraries that depend on other libraries (see Inter-library dependencies).

‘-o output-file’

Create output-file from the specified objects and libraries.

‘-release release’

Specify that the library was generated by release release of your package, so that users can easily tell which versions are newer than others. Be warned that no two releases of your package will be binary compatible if you use this flag. If you want binary compatibility, use the ‘-version-info’ flag (see Library interface versions).

‘-rpath libdir’

If output-file is a library, it will eventually be installed in libdir.

‘-static’

If output-file is a program, then do not link it against any uninstalled shared libtool libraries. If output-file is a library, then only create a static library.

‘-version-info current[:revision[:age]]’

If output-file is a libtool library, use interface version information current, revision, and age to build it (see Library interface versions). Do not use this flag to specify package release information, rather see the ‘-release’ flag.

If the output-file ends in ‘.la’, then a libtool library is created, which must be built only from library objects (‘.lo’ files). The ‘-rpath’ option is required. In the current implementation, libtool libraries may not depend on other uninstalled libtool libraries (see Inter-library dependencies).

If the output-file ends in ‘.a’, then a standard library is created using ar and possibly ranlib.

If output-file ends in ‘.o’ or ‘.lo’, then a reloadable object file is created from the input files (generally using ‘ld -r’). This method is often called partial linking.

Otherwise, an executable program is created.

4.3 Execute mode

For ‘execute’ mode, the library path is automatically set, then a program is executed.

The first of the mode-args is treated as a program name, with the rest as arguments to that program.

The following components of mode-args are treated specially:

‘-dlopen file’

Add the directory containing file to the library path.

This mode sets the library path environment variable according to any ‘-dlopen’ flags.

If any of the args are libtool executable wrappers, then they are translated into the name of their corresponding uninstalled binary, and any of their required library directories are added to the library path.

4.4 Install mode

In ‘install’ mode, libtool interprets mode-args as an installation command beginning with cp, or a BSD-compatible install program.

The rest of the mode-args are interpreted as arguments to that command.

The command is run, and any necessary unprivileged post-installation commands are also completed.

4.5 Finish mode

‘finish’ mode helps system administrators install libtool libraries so that they can be located and linked into user programs.

Each mode-arg is interpreted as the name of a library directory. Running this command may require superuser privileges, so the ‘--dry-run’ option may be useful.

4.6 Uninstall mode

This mode deletes installed libraries (and other files).

The first mode-arg is the name of the program to use to delete files (typically /bin/rm).

The remaining mode-args are either flags for the deletion program (beginning with a ‘-’), or the names of files to delete.

5.1 Writing Makefile rules for libtool

Libtool is fully integrated with Automake (see Introduction in The Automake Manual), starting with Automake version 1.2.

If you want to use libtool in a regular Makefile (or Makefile.in), you are on your own. If you’re not using Automake 1.2, and you don’t know how to incorporate libtool into your package you need to do one of the following:

1. Download Automake (version 1.2 or later) from your nearest GNU mirror, install it, and start using it.
2. Learn how to write Makefile rules by hand. They’re sometimes complex, but if you’re clever enough to write rules for compiling your old libraries, then you should be able to figure out new rules for libtool libraries (hint: examine the Makefile.in in the demo subdirectory of the libtool distribution… note especially that it was automatically generated from the Makefile.am by Automake).

5.2 Using Automake with libtool

Libtool library support is implemented under the ‘LTLIBRARIES’ primary.

Here are some samples from the Automake Makefile.am in the libtool distribution’s demo subdirectory.

First, to link a program against a libtool library, just use the ‘program_LDADD’ variable:

bin_PROGRAMS = hell hell.debug

# Build hell from main.c and libhello.la
hell_SOURCES = main.c
hell_LDADD = libhello.la

# Create an easier-to-debug version of hell.
hell_debug_SOURCES = main.c
hell_debug_LDADD = libhello.la
hell_debug_LDFLAGS = -static

You may use the ‘program_LDFLAGS’ variable to stuff in any flags you want to pass to libtool while linking ‘program’ (such as ‘-static’ to avoid linking uninstalled shared libtool libraries).

Building a libtool library is almost as trivial… note the use of ‘libhello_la_LDFLAGS’ to pass the ‘-version-info’ (see Library interface versions) option to libtool:

# Build a libtool library, libhello.la for installation in libdir.
lib_LTLIBRARIES = libhello.la
libhello_la_SOURCES = hello.c foo.c
libhello_la_LDFLAGS = -version-info 3:12:1

The ‘-rpath’ option is passed automatically by Automake, so you should not specify it.

See The Automake Manual in The Automake Manual, for more information.

5.3 Configuring libtool

Libtool requires intimate knowledge of your compiler suite and operating system in order to be able to create shared libraries and link against them properly. When you install the libtool distribution, a system-specific libtool script is installed into your binary directory.

However, when you distribute libtool with your own packages (see Including libtool with your package), you do not always know which compiler suite and operating system are used to compile your package.

For this reason, libtool must be configured before it can be used. This idea should be familiar to anybody who has used a GNU configure script. configure runs a number of tests for system features, then generates the Makefiles (and possibly a config.h header file), after which you can run make and build the package.

Libtool has its own equivalent to the configure script, ltconfig.

• Invoking ltconfig
• Using ltconfig
• The AM_PROG_LIBTOOL macro

ltconfig [option]… ltmain [host]

burger$ ./ltconfig ltmain.sh
checking host system type... i386-unknown-netbsd1.2
checking for ranlib... ranlib
checking for gcc... gcc
checking whether we are using GNU C... yes
checking for gcc option to produce PIC... -fPIC -DPIC
checking for gcc option to statically link programs... -static
checking if ld is GNU ld... no
checking if ld supports shared libraries... yes
checking dynamic linker characteristics... netbsd1.2 ld.so
checking if libtool supports shared libraries... yes
checking whether to build shared libraries... yes
creating libtool
burger$

burger$ export PATH=/local/i486-gnu/bin:$PATH
burger$ ./ltconfig ltmain.sh i486-gnu0.1
checking host system type... i486-unknown-gnu0.1
checking for ranlib... ranlib
checking for gcc... gcc
checking whether we are using GNU C... yes
checking for gcc option to produce PIC... -fPIC -DPIC
checking for gcc option to statically link programs... -static
checking if ld is GNU ld... yes
checking if GNU ld supports shared libraries... yes
checking dynamic linker characteristics... gnu0.1 ld.so
checking if libtool supports shared libraries... yes
checking whether to build shared libraries... yes
creating libtool
burger$

5.4 Including libtool with your package

In order to use libtool, you need to include the following files with your package:

config.guess ¶

Attempt to guess a canonical system name.

config.sub ¶

Canonical system name validation subroutine script.

ltconfig

Generate a libtool script for a given system.

ltmain.sh ¶

A generic script implementing basic libtool functionality.

Note that the libtool script itself should not be included with your package. See Configuring libtool.

You should use the libtoolize program, rather than manually copying these files into your package.

• Invoking libtoolize
• Autoconf ‘.o’ macros

libtoolize [option]…

LTLIBOBJS=`echo "$LIBOBJS" | sed 's/\.o/.lo/g'`
AC_SUBST(LTLIBOBJS)
LTALLOCA=`echo "$ALLOCA" | sed 's/\.o/.lo/g'`
AC_SUBST(LTALLOCA)
AC_OUTPUT(…)

5.5 Static-only libraries

When you are developing a package, it is often worthwhile to configure your package with the ‘--disable-shared’ flag, or to override the defaults for AM_PROG_LIBTOOL by using the AM_DISABLE_SHARED Autoconf macro (see The AM_PROG_LIBTOOL macro). This prevents libtool from building shared libraries, which has several advantages:

• compilation is twice as fast, which can speed up your development cycle
• debugging is easier because you don’t need to deal with any complexities added by shared libraries
• you can see how libtool behaves on static-only platforms

You may want to put a small note in your package README to let other developers know that ‘--disable-shared’ can save them time. The following example note is taken from the GIMP⁵ distribution README:

The GIMP uses GNU Libtool in order to build shared libraries on a
variety of systems. While this is very nice for making usable
binaries, it can be a pain when trying to debug a program. For that
reason, compilation of shared libraries can be turned off by
specifying the ‘--disable-shared’ option to configure.

6.1 What are library interfaces?

Interfaces for libraries may be any of the following (and more):

• global variables: both names and types
• global functions: argument types and number, return types, and function names
• standard input, standard output, standard error, and file formats
• sockets, pipes, and other inter-process communication protocol formats

Note that static functions do not count as interfaces, because they are not directly available to the user of the library.

6.2 Libtool’s versioning system

Libtool has its own formal versioning system. It is not as flexible as some, but it is definitely the simplest of the more powerful versioning systems.

Think of a library as exporting several sets of interfaces, arbitrarily represented by integers. When a program is linked against a library, it may use any subset of those interfaces.

Libtool’s description of the interfaces that a program uses is very simple: it encodes the least and the greatest interface numbers in the resulting binary (first-interface, last-interface).

The dynamic linker is guaranteed that if a library supports every interface number between first-interface and last-interface, then the program can be relinked against that library.

Note that this can cause problems because libtool’s compatibility requirements are actually stricter than is necessary.

Say libhello supports interfaces 5, 16, 17, 18, and 19, and that libtool is used to link test against libhello.

Libtool encodes the numbers 5 and 19 in test, and the dynamic linker will only link test against libraries that support every interface between 5 and 19. So, the dynamic linker refuses to link test against libhello!

In order to eliminate this problem, libtool only allows libraries to declare consecutive interface numbers. So, libhello can declare at most that it supports interfaces 16 through 19. Then, the dynamic linker will link test against libhello.

So, libtool library versions are described by three integers:

current

The most recent interface number that this library implements.

revision

The implementation number of the current interface.

age

The difference between the newest and oldest interfaces that this library implements. In other words, the library implements all the interface numbers in the range from number current - age to current.

If two libraries have identical current and age numbers, then the dynamic linker chooses the library with the greater revision number.

6.3 Updating library version information

If you want to use libtool’s versioning system, then you must specify the version information to libtool using the ‘-version-info’ flag during link mode (see Link mode).

This flag accepts an argument of the form ‘current[:revision[:age]]’. So, passing ‘-version-info 3:12:1’ sets current to 3, revision to 12, and age to 1.

If either revision or age are omitted, they default to 0. Also note that age must be less than or equal to the current interface number.

Here are a set of rules to help you update your library version information:

1. Start with version information of ‘0:0:0’ for each libtool library.
2. Update the version information only immediately before a public release of your software. More frequent updates are unnecessary, and only guarantee that the current interface number gets larger faster.
3. If the library source code has changed at all since the last update, then increment revision (‘c:r:a’ becomes ‘c:r+1:a’).
4. If any interfaces have been added, removed, or changed since the last update, increment current, and set revision to 0.
5. If any interfaces have been added since the last public release, then increment age.
6. If any interfaces have been removed since the last public release, then set age to 0.

Never try to set the interface numbers so that they correspond to the release number of your package. This is an abuse that only fosters misunderstanding of the purpose of library versions. Instead, use the ‘-release’ flag (see Managing release information), but be warned that every release of your package will not be binary compatibility with any other release.

6.4 Managing release information

Often, people want to encode the name of the package release into the shared library so that it is obvious to the user which package their programs are linked against. This convention is used especially on Linux:

trick$ ls /usr/lib/libbfd*
/usr/lib/libbfd.a	    /usr/lib/libbfd.so.2.7.0.2
/usr/lib/libbfd.so
trick$

On ‘trick’, /usr/lib/libbfd.so is just a symbolic link to /usr/lib/libbfd.so.2.7.0.2, which was distributed as a part of ‘binutils-2.7.0.2’.

Unfortunately, this convention conflicts directly with libtool’s idea of library interface versions, because the library interface rarely changes at the same time that the release number does, and the library suffix is never the same across all platforms.

So, in order to accomodate both views, you can use the ‘-release’ flag in order to set release information for libraries which you do not want to use ‘-version-info’. For the libbfd example, the next release which uses libtool should be built with ‘-release 2.9.0’, which will produce the following files on Linux:

trick$ ls /usr/lib/libbfd*
/usr/lib/libbfd-2.9.0.so.0         /usr/lib/libbfd.so
/usr/lib/libbfd-2.9.0.so.0.0.0     /usr/lib/libbfd.a
trick$

In this case, /usr/lib/libbfd.so is a symbolic link to /usr/lib/libbfd-2.9.0.so.0.0.0. This makes it obvious that the user is dealing with ‘binutils-2.9.0’, without compromising libtool’s idea of interface versions.

Note that this option actually causes a modification of the library name, so do not use it if unless you want to break binary compatibility with any past library releases. In general, you should only use ‘-release’ for libraries whose interfaces change very frequently.

7.1 Writing C header files

Writing portable C header files can be difficult, since they may be read by different types of compilers:

C++ compilers

C++ compilers require that functions be declared with full prototypes, since C++ is more strongly typed than C. C functions and variables also need to be declared with the extern "C" directive, so that the names aren’t mangled. See Writing libraries for C++, for other issues relevant to using C++ with libtool.

ANSI C compilers

ANSI C compilers are not as strict as C++ compilers, but functions should be prototyped to avoid unnecessary warnings when the header file is #included.

non-ANSI C compilers

Non-ANSI compilers will report errors if functions are prototyped.

These complications mean that your library interface headers must use some C preprocessor magic in order to be usable by each of the above compilers.

foo.h in the demo subdirectory of the libtool distribution serves as an example for how to write a header file that can be safely installed in a system directory.

Here are the relevant portions of that file:

/* __BEGIN_DECLS should be used at the beginning of your declarations,
   so that C++ compilers don't mangle their names.  Use __END_DECLS at
   the end of C declarations. */
#undef __BEGIN_DECLS
#undef __END_DECLS
#ifdef __cplusplus
# define __BEGIN_DECLS extern "C" {
# define __END_DECLS }
#else
# define __BEGIN_DECLS /* empty */
# define __END_DECLS /* empty */
#endif

/* __P is a macro used to wrap function prototypes, so that compilers
   that don't understand ANSI C prototypes still work, and ANSI C
   compilers can issue warnings about type mismatches. */
#undef __P
#if defined (__STDC__) || defined (_AIX) \
        || (defined (__mips) && defined (_SYSTYPE_SVR4)) \
        || defined(WIN32) || defined(__cplusplus)
# define __P(protos) protos
#else
# define __P(protos) ()
#endif

These macros are used in foo.h as follows:

#ifndef _FOO_H_
#define _FOO_H_ 1

/* The above macro definitions. */
…

__BEGIN_DECLS
int foo __P((void));
int hello __P((void));
__END_DECLS

#endif /* !_FOO_H_ */

Note that the #ifndef _FOO_H_ prevents the body of foo.h from being read more than once in a given compilation.

Feel free to copy the definitions of __P, __BEGIN_DECLS, and __END_DECLS into your own headers. Then, you may use them to create header files that are valid for C++, ANSI, and non-ANSI compilers.

Do not be naive about writing portable code. Following the tips given above will help you miss the most obvious problems, but there are definitely other subtle portability issues. You may need to cope with some of the following issues:

• Pre-ANSI compilers do not always support the void * generic pointer type, and so need to use char * in its place.
• The const and signed keywords are not supported by some compilers, especially pre-ANSI compilers.
• The long double type is not supported by many compilers.

9.1 Building modules to dlopen

On some operating systems, a program symbol must be specially declared in order to be dynamically resolved with the dlsym(3) (or equivalent) function.

Libtool provides the ‘-export-dynamic’ link flag (see Link mode), which does this declaration. You need to use this flag if you are linking an application program that dlopens other modules or a libtool library that will also be dlopened.

For example, if we wanted to build a shared library, libhello, that would later be dlopened by an application, we would add ‘-export-dynamic’ to the other link flags:

burger$ libtool gcc -export-dynamic -o libhello.la foo.lo \
                hello.lo -rpath /usr/local/lib -lm
burger$

Another situation where you would use ‘-export-dynamic’ is if symbols from your executable are needed to satisfy unresolved references in a library you want to dlopen. In this case, you should use ‘-export-dynamic’ while linking the executable that calls dlopen:

burger$ libtool gcc -export-dynamic -o hell-dlopener main.o
burger$

9.2 Dlpreopening

Libtool provides special support for dlopening libtool object and libtool library files, so that their symbols can be resolved even on platforms without any dlopen(3) and dlsym(3) functions..

Consider the following alternative ways of loading code into your program, in order of increasing “laziness”:

1. Linking against object files that become part of the program executable, whether or not they are referenced. If an object file cannot be found, then the linker refuses to create the executable.
2. Declaring a static library to the linker, so that it is searched at link time in order to satisfy any undefined references in the above object files. If the static library cannot be found, then the linker refuses to link the executable.
3. Declaring a shared library to the runtime linker, so that it is searched at runtime in order to satisfy any undefined references in the above files. If the shared library cannot be found, then the dynamic linker aborts the program before it runs.
4. Dlopening a module, so that the application can resolve its own, dynamically-computed references. If there is an error opening the module, or the module is not found, then the application can recover without crashing.

Libtool emulates ‘-export-dynamic’ on static platforms by linking objects into the program at compile time, and creating data structures that represent the program’s symbol table.

In order to use this feature, you must declare the objects you want your application to dlopen by using the ‘-dlopen’ or ‘-dlpreopen’ flags when you link your program (see Link mode).

Structure: dld_symbol name address ¶

The name attribute is a 0-terminated character string of the symbol name, such as "fprintf". The address attribute is a generic pointer to the appropriate object, which is &fprintf in this example.

Variable: dld_symbol * dld_preloaded_symbols ¶

An array of dld_symbol structures, representing all the preloaded symbols linked into the program. The last element has a name of 0.

Variable: int dld_preloaded_symbol_count ¶

The number of elements in dld_preloaded_symbols, if it is sorted in ascending order by name. Otherwise, -1, to indicate that the application needs to sort and count dld_preloaded_symbols itself, or search it linearly.

Some compilers may allow identifiers which are not valid in ANSI C, such as dollar signs. Libtool only recognizes valid ANSI C symbols (an initial ASCII letter or underscore, followed by zero or more ASCII letters, digits, and underscores), so non-ANSI symbols will not appear in dld_preloaded_symbols.

9.3 Finding the correct name to dlopen

After a library has been linked with ‘-export-dynamic’, it can be dlopened. Unfortunately, because of the variation in library names, your package needs to determine the correct file to dlopen.

The most straightforward and flexible implementation is to determine the name at runtime, by finding the installed ‘.la’ file, and searching it for the following lines:

# The name that we can dlopen(3).
dlname='dlname'

If dlname is empty, then the library cannot be dlopened. Otherwise, it gives the dlname of the library. So, if the library was installed as /usr/local/lib/libhello.la, and the dlname was libhello.so.3, then /usr/local/lib/libhello.so.3 should be dlopened.

If your program uses this approach, then it should search the directories listed in the LD_LIBRARY_PATH⁸ environment variable, as well as the directory where libraries will eventually be installed. Searching this variable (or equivalent) will guarantee that your program can find its dlopened modules, even before installation, provided you have linked them using libtool.

9.4 Unresolved dlopen issues

The following problems are not solved by using libtool’s dlopen support:

• Dlopen functions are generally only available on shared library platforms. If you want your package to be portable to static platforms, you have to develop your own alternatives to dlopening dynamic code. Most reasonable solutions involve writing wrapper functions for the dlopen(3) family, which do package-specific tricks when dlopening is unsupported or not available on a given platform.
• There are major differences in implementations of the dlopen(3) family of functions. Some platforms do not even use the same function names (notably HP-UX, with its ‘shl_load(3)’ family).
• The application developer must write a custom search function in order to discover the correct module filename to supply to dlopen(3).

Each of these limitations will be addressed in GNU DLD 4.⁹

10.1 Writing libraries for C++

Creating libraries of C++ code is a fairly straightforward process, and differs from C code in only two ways:

1. Because of name mangling, C++ libraries are only usable by the C++ compiler that created them. This decision was made by the designers of C++ in order to protect users from conflicting implementations of features such as constructors, exception handling, and RTTI.
2. On some systems, notably SunOS 4, the dynamic linker does not call non-constant initializers. This can lead to hard-to-pinpoint bugs in your library. GCC 2.7 and later versions work around this problem, but previous versions and other compilers do not.

This second issue is very complex. Basically, you should avoid any global or static variable initializations that would cause an “initializer element is not constant” error if you compiled them with a standard C compiler.

There are other ways of working around this problem, but they are beyond the scope of this manual.

11.1 The libtool test suite

Libtool comes with its own set of programs that test its capabilities, and report obvious bugs in the libtool program. These tests, too, are constantly evolving, based on past problems with libtool, and known deficiencies in other operating systems.

As described in the INSTALL file, you may run make check after you have built libtool (possibly before you install it) in order to make sure that it meets basic functional requirements.

• Description of test suite
• When tests fail

11.2 Reporting bugs

If you think you have discovered a bug in libtool, you should think twice: the libtool maintainer is notorious for passing the buck (or maybe that should be “passing the bug”). Libtool was invented to fix known deficiencies in shared library implementations, so, in a way, most of the bugs in libtool are actually bugs in other operating systems. However, the libtool maintainer would definitely be happy to add support for somebody else’s buggy operating system. [I wish there was a good way to do winking smiley-faces in texinfo.]

Genuine bugs in libtool include problems with shell script portability, documentation errors, and failures in the test suite (see The libtool test suite).

First, check the documentation and help screens to make sure that the behaviour you think is a problem is not already mentioned as a feature.

Then, you should read the Emacs guide to reporting bugs (see Reporting Bugs in The Emacs Manual). Some of the details listed there are specific to Emacs, but the principle behind them is a general one.

Finally, send a bug report to the libtool mailing list <bug-libtool@gnu.org> with any appropriate facts, such as test suite output (see When tests fail), all the details needed to reproduce the bug, and a brief description of why you think the behaviour is a bug. Be sure to include the word “libtool” in the subject line, as well as the version number you are using (which can be found by typing ltconfig --version).

Please include the generated libtool script with your bug report, so that I can see what values ltconfig guessed for your system.

12.1 Porting libtool to new systems

To port libtool to a new system, you’ll generally need the following information:

man pages for ld(1) and cc(1)

These generally describe what flags are used to generate PIC, to create shared libraries, and to link against only static libraries. You may need to follow some cross references to find the information that is required.

man pages for ld.so(8), rtld(8), or equivalent

These are a valuable resource for understanding how shared libraries are loaded on the system.

man page for ldconfig(8), or equivalent

This page usually describes how to install shared libraries.

output from ls -l /lib /usr/lib

This shows the naming convention for shared libraries on the system, including which names should be symbolic links.

any additional documentation

Some systems have special documentation on how to build and install shared libraries.

12.2 Tested platforms

This table describes when libtool was last known to be tested on platforms where it claims to support shared libraries:

--------------------------------------------------------
canonical host name          compiler  libtool   results
                                       release
--------------------------------------------------------
alpha-dec-osf3.2             cc        0.8       ok
alpha-dec-osf3.2             gcc       0.8       ok
alpha-dec-osf4.0             cc        1.0f      ok
alpha-dec-osf4.0             gcc       1.0f      ok
alpha-unknown-linux          gcc       0.9h      ok
hppa1.1-hp-hpux9.07          cc        1.0f      ok
hppa1.1-hp-hpux9.07          gcc       1.0f      ok
hppa1.1-hp-hpux10.10         cc        0.9h      ok
hppa1.1-hp-hpux10.10         gcc       0.9h      ok
i386-unknown-freebsd2.1.5    gcc       0.5       ok
i386-unknown-gnu0.0          gcc       0.5       ok
i386-unknown-netbsd1.2       gcc       0.9g      ok
i586-pc-linux-gnulibc1       gcc       1.0i      ok
i586-pc-linux-gnu            gcc       1.0i      ok
mips-sgi-irix5.2             gcc       1.0i      ok
mips-sgi-irix5.3             cc        0.8       ok
mips-sgi-irix5.3             gcc       0.8       ok
mips-sgi-irix6.2             cc        0.9       ok
mips-sgi-irix6.3             cc        1.0f      ok
mips-sgi-irix6.3             gcc       1.0i      ok
mips-sgi-irix6.3             irix5-gcc 1.0f      ok
mipsel-unknown-openbsd2.1    gcc       1.0       ok
powerpc-ibm-aix4.1.4.0       xlc       1.0i      ok
powerpc-ibm-aix4.1.4.0       gcc       1.0       ok
rs6000-ibm-aix3.2.5          xlc       1.0i      ok
rs6000-ibm-aix3.2.5          gcc       1.0i      ok*
sparc-sun-linux2.1.23        gcc       0.9h      ok
sparc-sun-sunos4.1.3         gcc       1.0i      ok
sparc-sun-sunos4.1.4         cc        1.0f      ok
sparc-sun-sunos4.1.4         gcc       1.0f      ok
sparc-sun-solaris2.4         cc        1.0a      ok
sparc-sun-solaris2.4         gcc       1.0a      ok
sparc-sun-solaris2.5         cc        1.0f      ok
sparc-sun-solaris2.5         gcc       1.0i      ok
sparc-sun-solaris2.6         gcc       1.0i      ok
--------------------------------------------------------

* Some versions of GCC's collect2 linker program cannot link trivial
static binaries on AIX 3.  For these configurations, libtool's `-static'
flag has no effect.

12.3 Platform quirks

This section is dedicated to the sanity of the libtool maintainer. It describes the programs that libtool uses, how they vary from system to system, and how to test for them.

Because libtool is a shell script, it is very difficult to understand just by reading it from top to bottom. This section helps show why libtool does things a certain way. After reading it, then reading the scripts themselves, you should have a better sense of how to improve libtool, or write your own.

• References
• Compilers
• Reloadable objects
• Archivers

12.4 libtool script contents

The libtool script is generated by ltconfig (see Configuring libtool). Ever since libtool version 0.7, this script simply sets shell variables, then sources the libtool backend, ltmain.sh.

Here is a listing of each of these variables, and how they are used within ltmain.sh:

Variable: AR ¶

The name of the system library archiver.

Variable: CC ¶

The name of the C compiler used to configure libtool.

Variable: LD ¶

The name of the linker that libtool should use internally for reloadable linking and possibly shared libraries.

Variable: LTCONFIG_VERSION ¶

This is set to the version number of the ltconfig script, to prevent mismatches between the configuration information in libtool, and how that information is used in ltmain.sh.

Variable: NM ¶

The name of a BSD-compatible nm program, which produces listings of global symbols in one the following formats:

address C global-variable-name
address D global-variable-name
address T global-function-name

Variable: RANLIB ¶

Set to the name of the ranlib program, if any.

Variable: allow_undefined_flag ¶

The flag that is used by ‘archive_cmds’ in order to declare that there will be unresolved symbols in the resulting shared library. Empty, if no such flag is required. Set to ‘unsupported’ if there is no way to generate a shared library with references to symbols that aren’t defined in that library.

Variable: archive_cmds ¶

Variable: old_archive_cmds ¶

Commands used to create shared and static libraries, respectively.

Variable: build_libtool_libs ¶

Whether libtool should build shared libraries on this system. Set to ‘yes’ or ‘no’.

Variable: build_old_libs ¶

Whether libtool should build static libraries on this system. Set to ‘yes’ or ‘no’.

Variable: echo ¶

An echo(1) program which does not interpret backslashes as an escape character.

Variable: export_dynamic_flag_spec ¶

Compiler link flag that allows a dlopened shared library to reference symbols that are defined in the program.

Variable: finish_cmds ¶

Commands to tell the dynamic linker how to find shared libraries in a specific directory.

Variable: finish_eval ¶

Same as finish_cmds, except the commands are not displayed.

Variable: global_symbol_pipe ¶

A pipeline that takes the output of NM, and produces a listing of raw symbols followed by their C names. For example:

$ $NM | $global_symbol_pipe
symbol1 C-symbol1
symbol2 C-symbol2
symbol3 C-symbol3
…
$

Variable: hardcode_action ¶

Either ‘immediate’ or ‘relink’, depending on whether shared library paths can be hardcoded into executables before they are installed, or if they need to be relinked.

Variable: hardcode_direct ¶

Set to ‘yes’ or ‘no’, depending on whether the linker hardcodes directories if a library is directly specified on the command line (such as ‘dir/libname.a’).

Variable: hardcode_libdir_flag_spec ¶

Flag to hardcode a libdir variable into a binary, so that the dynamic linker searches libdir for shared libraries at runtime.

Variable: hardcode_libdir_separator ¶

If the compiler only accepts a single hardcode_libdir_flag, then this variable contains the string that should separate multiple arguments to that flag.

Variable: hardcode_minus_L ¶

Set to ‘yes’ or ‘no’, depending on whether the linker hardcodes directories specified by ‘-L’ flags into the resulting executable.

Variable: hardcode_shlibpath_var ¶

Set to ‘yes’ or ‘no’, depending on whether the linker hardcodes directories by writing the contents of ‘$shlibpath_var’ into the resulting executable. Set to ‘unsupported’ if directories specified by ‘$shlibpath_var’ are searched at run time, but not at link time.

Variable: host ¶

Variable: host_alias ¶

For information purposes, set to the specified and canonical names of the system that libtool was configured for.

Variable: libname_spec ¶

The format of a library name prefix. On all Unix systems, static libraries are called ‘libname.a’, but on some systems (such as OS/2 or MS-DOS), the library is just called ‘name.a’.

Variable: library_names_spec ¶

A list of shared library names. The first is the name of the file, the rest are symbolic links to the file. The name in the list is the file name that the linker finds when given ‘-lname’.

Variable: link_static_flag ¶

Linker flag (passed through the C compiler) used to prevent dynamic linking.

Variable: no_builtin_flag ¶

Compiler flag to disable builtin functions that conflict with declaring external global symbols as char.

Variable: no_undefined_flag ¶

The flag that is used by ‘archive_cmds’ in order to declare that there will be no unresolved symbols in the resulting shared library. Empty, if no such flag is required.

Variable: pic_flag ¶

Any additional compiler flags for building library object files.

Variable: postinstall_cmds ¶

Variable: old_postinstall_cmds ¶

Commands run after installing a shared or static library, respectively.

Variable: reload_cmds ¶

Variable: reload_flag ¶

Commands to create a reloadable object.

Variable: runpath_var ¶

The environment variable that tells the linker which directories to hardcode in the resulting executable.

Variable: shlibpath_var ¶

The environment variable that tells the dynamic linker where to find shared libraries.

Variable: soname_spec ¶

The name coded into shared libraries, if different from the real name of the file.

Variable: version_type ¶

The library version numbering type. One of ‘libtool’, ‘linux’, ‘osf’, ‘sunos’, or ‘none’.

Variable: wl ¶

The C compiler flag that allows libtool to pass a flag directly to the linker. Used as: ‘${wl}some-flag’.

Variables ending in ‘_cmds’ or ‘_eval’ contain a semicolon-separated list of commands that are evaled one after another. If any of the commands return a nonzero exit status, libtool generally exits with an error message.

Variables ending in ‘_spec’ are evaled before being used by libtool.