How To Write Shared Libraries

Ulrich Drepper
[email protected]

December 10, 2011

Abstract
Today, shared libraries are ubiquitous. Developers use them for multiple reasons and create
them just as they would create application code. This is a problem, though, since on many
platforms some additional techniques must be applied even to generate decent code. Even more
knowledge is needed to generate optimized code. This paper introduces the required rules and
techniques. In addition, it introduces the concept of ABI (Application Binary Interface) stability
and shows how to manage it.

1 Preface 1.1 A Little Bit of History

The binary format used initially for Linux was an a.out

For a long time, programmers collected commonly used variant. When introducing shared libraries certain design
code in libraries so that code could be reused. This saves decisions had to be made to work in the limitations of
development time and reduces errors since reused code a.out. The main accepted limitation was that no reloca-
only has to be debugged once. With systems running tions are performed at the time of loading and afterward.
dozens or hundreds of processes at the same time reuse The shared libraries have to exist in the form they are
of the code at link-time solves only part of the problem. used at run-time on disk. This imposes a major restric-
Many processes will use the same pieces of code which tion on the way shared libraries are built and used: every
they import for libraries. With the memory management shared library must have a fixed load address; otherwise it
systems in modern operating systems it is also possible would not be possible to generate shared libraries which
to share the code at run-time. This is done by loading the do not have to be relocated.
code into physical memory only once and reusing it in
multiple processes via virtual memory. Libraries of this The fixed load addresses had to be assigned and this has
kind are called shared libraries. to happen without overlaps and conflicts and with some
future safety by allowing growth of the shared library.
The concept is not very new. Operating system designers It is therefore necessary to have a central authority for
implemented extensions to their system using the infras- the assignment of address ranges which in itself is a ma-
tructure they used before. The extension to the OS could jor problem. But it gets worse: given a Linux system
be done transparently for the user. But the parts the user of today with many hundred of DSOs (Dynamic Shared
directly has to deal with created initially problems. Objects) the address space and the virtual memory avail-
able to the application gets severely fragmented. This
The main aspect is the binary format. This is the for- would limit the size of memory blocks which can be dy-
mat which is used to describe the application code. Long namically allocated which would create unsurmountable
gone are the days that it was sufficient to provide a mem- problems for some applications. It would even have hap-
ory dump. Multi-process systems need to identify differ- pened by today that the assignment authority ran out of
ent parts of the file containing the program such as the address ranges to assign, at least on 32-bit machines.
text, data, and debug information parts. For this, binary
formats were introduced early on. Commonly used in the We still have not covered all the drawbacks of the a.out
early Unix-days were formats such as a.out or COFF. shared libraries. Since the applications using shared li-
These binary formats were not designed with shared li- braries should not have to be relinked after changing a
braries in mind and this clearly shows. shared library it uses, the entry points, i.e., the function
and variable addresses, must not change. This can only
be guaranteed if the entry points are kept separate from
the actual code since otherwise limits on the size of a
function would be hard-coded. A table of function stubs
Copyright c 2002-2010, 2011 Ulrich Drepper
which call the actual implementation was the solution
All rights reserved. No redistribution allowed. used on Linux. The static linker got the address of each

1
function stub from a special file (with the filename ex- be followed to get optimal results. Explaining these rules
tension .sa). At run-time a file ending in .so.X.Y.Z will be the topic of a large portion of this paper.
was used and it had to correspond to the used .sa file.
This in turn requires that an allocated entry in the stub Not all uses of DSOs are for the purpose of saving re-
table always had to be used for the same function. The sources. DSOs are today also often used as a way to
allocation of the table had to be carefully taken care of. structure programs. Different parts of the program are
Introducing a new interface meant appending to the ta- put into separate DSOs. This can be a very powerful tool,
ble. It was never possible to retire a table entry. To avoid especially in the development phase. Instead of relink-
using an old shared library with a program linked with a ing the entire program it is only necessary to relink the
newer version, some record had to be kept in the applica- DSO(s) which changed. This is often much faster.
tion: the X and Y parts of the name of the .so.X.Y.Z
suffix was recorded and the dynamic linker made sure Some projects decide to keep many separate DSOs even
minimum requirements were met. in the deployment phase even though the DSOs are not
reused in other programs. In many situations it is cer-
The benefits of the scheme are that the resulting program tainly a useful thing to do: DSOs can be updated indi-
runs very fast. Calling a function in such a shared li- vidually, reducing the amount of data which has to be
braries is very efficient even for the first call. It can transported. But the number of DSOs must be kept to a
be implemented with only two absolute jumps: the first reasonable level. Not all programs do this, though, and
from the user code to the stub, and the second from the we will see later on why this can be a problem.
stub to the actual code of the function. This is probably
faster than any other shared library implementation, but Before we can start discussing all this some understand-
its speed comes at too high a price: ing of ELF and its implementation is needed.

1.3 How Is ELF Implemented?

1. a central assignment of address ranges is needed;
2. collisions are possible (likely) with catastrophic re-
The handling of a statically linked application is very
sults;
simple. Such an application has a fixed load address
3. the address space gets severely fragmented. which the kernel knows. The load process consists sim-
ply of making the binary available in the appropriate ad-
dress space of a newly created process and transferring
For all these reasons and more, Linux converted early on control to the entry point of the application. Everything
to using ELF (Executable Linkage Format) as the binary else was done by the static linker when creating the exe-
format. The ELF format is defined by the generic spec- cutable.
ification (gABI) to which processor-specific extensions
(psABI) are added. As it turns out the amortized cost of Dynamically linked binaries, in contrast, are not com-
function calls is almost the same as for a.out but the plete when they are loaded from disk. It is therefore
restrictions are gone. not possible for the kernel to immediately transfer con-
trol to the application. Instead some other helper pro-
1.2 The Move To ELF gram, which obviously has to be complete, is loaded as
well. This helper program is the dynamic linker. The task
of the dynamic linker is it, to complete the dynamically
For programmers the main advantage of the switch to linked application by loading the DSOs it needs (the de-
ELF was that creating ELF shared libraries, or in ELF- pendencies) and to perform the relocations. Then finally
speak DSOs, becomes very easy. The only difference be- control can be transferred to the program.
tween generating an application and a DSO is in the final
link command line. One additional option (--shared in This is not the last task for the dynamic linker in most
the case of GNU ld) tells the linker to generate a DSO cases, though. ELF allows the relocations associated with
instead of an application, the latter being the default. In a symbol to be delayed until the symbol is needed. This
fact, DSOs are little more than a special kind of binary; lazy relocation scheme is optional, and optimizations dis-
the difference is that they have no fixed load address and cussed below for relocations performed at startup imme-
hence require the dynamic linker to actually become ex- diately effect the lazy relocations as well. So we ignore
ecutable. With Position Independent Executable (PIEs) in the following everything after the startup is finished.
the difference shrinks even more.
1.4 Startup: In The Kernel
This, together with the introduction of GNU Libtool which
will be described later, has led to the wide adoption of
DSOs by programmers. Proper use of DSOs can help Starting execution of a program begins in the kernel, nor-
save large amounts of resources. But some rules must be mally in the execve system call. The currently executed
followed to get any benefits, and some more rules have to code is replaced with a new program. This means the ad-

2 Version 4.1.2 How To Write Shared Libraries

 typedef struct typedef struct
{ {
Elf32 Word p type; Elf64 Word p type;
Elf32 Off p offset; Elf64 Word p flags;
Elf32 Addr p vaddr; Elf64 Off p offset;
Elf32 Addr p paddr; Elf64 Addr p vaddr;
Elf32 Word p filesz; Elf64 Addr p paddr;
Elf32 Word p memsz; Elf64 Xword p filesz;
Elf32 Word p flags; Elf64 Xword p memsz;
Elf32 Word p align; Elf64 Xword p align;
} Elf32 Phdr; } Elf64 Phdr;

Figure 1: ELF Program Header C Data Structure

dress space content is replaced by the content of the file contains the size of each entry. This last value is useful
containing the program. This does not happen by sim- only as a run-time consistency check for the binary.
ply mapping (using mmap) the content of the file. ELF
files are structured and there are normally at least three The different segments are represented by the program
different kinds of regions in the file: header entries with the PT LOAD value in the p type field.
The p offset and p filesz fields specify where in the
file the segment starts and how long it is. The p vaddr
• Code which is executed; this region is normally not and p memsz fields specify where the segment is located
writable; in the the process’ virtual address space and how large the
memory region is. The value of the p vaddr field itself
• Data which is modified; this region is normally not
is not necessarily required to be the final load address.
executable;
DSOs can be loaded at arbitrary addresses in the virtual
• Data which is not used at run-time; since not needed address space. But the relative position of the segments
it should not be loaded at startup. is important. For pre-linked DSOs the actual value of the
p vaddr field is meaningful: it specifies the address for
which the DSO was pre-linked. But even this does not
Modern operating systems and processors can protect mem- mean the dynamic linker cannot ignore this information
ory regions to allow and disallow reading, writing, and if necessary.
executing separately for each page of memory.1 It is
preferable to mark as many pages as possible not writable The size in the file can be smaller than the address space
since this means that the pages can be shared between it takes up in memory. The first p filesz bytes of the
processes which use the same application or DSO the memory region are initialized from the data of the seg-
page is from. Write protection also helps to detect and ment in the file, the difference is initialized with zero.
prevent unintentional or malignant modifications of data This can be used to handle BSS sections2 , sections for
or even code. uninitialized variables which are according to the C stan-
dard initialized with zero. Handling uninitialized vari-
For the kernel to find the different regions, or segments ables this way has the advantage that the file size can be
in ELF-speak, and their access permissions, the ELF file reduced since no initialization value has to be stored, no
format defines a table which contains just this informa- data has to be copied from disc to memory, and the mem-
tion, among other things. The ELF Program Header ta- ory provided by the OS via the mmap interface is already
ble, as it is called, must be present in every executable initialized with zero.
and DSO. It is represented by the C types Elf32 Phdr
and Elf64 Phdr which are defined as can be seen in fig- The p flags finally tells the kernel what permissions to
ure 1. use for the memory pages. This field is a bitmap with the
bits given in the following table being defined. The flags
To locate the program header data structure another data are directly mapped to the flags mmap understands.
structure is needed, the ELF Header. The ELF header is
the only data structure which has a fixed place in the file,
starting at offset zero. Its C data structure can be seen
in figure 2. The e phoff field specifies where, counting
from the beginning of the file, the program header table
starts. The e phnum field contains the number of entries
in the program header table and the e phentsize field
2 A BSS section contains only NUL bytes. Therefore they do not
1A memory page is the smallest entity the memory subsystem of have to be represented in the file on the storage medium. The loader
the OS operates on. The size of a page can vary between different just has to know the size so that it can allocate memory large enough
architectures and even within systems using the same architecture. and fill it with NUL

Ulrich Drepper Version 4.1.2 3

 typedef struct typedef struct
{ {
unsigned char e ident[EI NIDENT]; unsigned char e ident[EI NIDENT];
Elf32 Half e type; Elf64 Half e type;
Elf32 Half e machine; Elf64 Half e machine;
Elf32 Word e version; Elf64 Word e version;
Elf32 Addr e entry; Elf64 Addr e entry;
Elf32 Off e phoff; Elf64 Off e phoff;
Elf32 Off e shoff; Elf64 Off e shoff;
Elf32 Word e flags; Elf64 Word e flags;
Elf32 Half e ehsize; Elf64 Half e ehsize;
Elf32 Half e phentsize; Elf64 Half e phentsize;
Elf32 Half e phnum; Elf64 Half e phnum;
Elf32 Half e shentsize; Elf64 Half e shentsize;
Elf32 Half e shnum; Elf64 Half e shnum;
Elf32 Half e shstrndx; Elf64 Half e shstrndx;
} Elf32 Ehdr; } Elf64 Ehdr;

Figure 2: ELF Header C Data Structure

p flags Value mmap flag Description to once the application is complete. For this a structured
PF X 1 PROT EXEC Execute Permission way exists. The kernel puts an array of tag-value pairs on
PF W 2 PROT WRITE Write Permission the stack of the new process. This auxiliary vector con-
PF R 4 PROT READ Read Permission tains beside the two aforementioned values several more
values which allow the dynamic linker to avoid several
system calls. The elf.h header file defines a number of
constants with a AT prefix. These are the tags for the
After mapping all the PT LOAD segments using the ap- entries in the auxiliary vector.
propriate permissions and the specified address, or after
freely allocating an address for dynamic objects which After setting up the auxiliary vector the kernel is finally
have no fixed load address, the next phase can start. The ready to transfer control to the dynamic linker in user
virtual address space of the dynamically linked executable mode. The entry point is defined in e entry field of the
itself is set up. But the binary is not complete. The kernel ELF header of the dynamic linker.
has to get the dynamic linker to do the rest and for this
the dynamic linker has to be loaded in the same way as
the executable itself (i.e., look for the loadable segments 1.5 Startup in the Dynamic Linker
in the program header). The difference is that the dy-
namic linker itself must be complete and should be freely
relocatable. The second phase of the program startup happens in the
dynamic linker. Its tasks include:
Which binary implements the dynamic linker is not hard-
coded in the kernel. Instead the program header of the
application contains an entry with the tag PT INTERP.
• Determine and load dependencies;
The p offset field of this entry contains the offset of
a NUL-terminated string which specifies the file name of
this file. The only requirement on the named file is that • Relocate the application and all dependencies;
its load address does not conflict with the load address of
any possible executable it might be used with. In gen-
• Initialize the application and dependencies in the
eral this means that the dynamic linker has no fixed load
correct order.
address and can be loaded anywhere; this is just what dy-
namic binaries allow.

Once the dynamic linker has also been mapped into the In the following we will discuss in more detail only the
memory of the to-be-started process we can start the dy- relocation handling. For the other two points the way
namic linker. Note it is not the entry point of the applica- for better performance is clear: have fewer dependen-
tion to which control is transfered to. Only the dynamic cies. Each participating object is initialized exactly once
linker is ready to run. Instead of calling the dynamic but some topological sorting has to happen. The identify
linker right away, one more step is performed. The dy- and load process also scales with the number dependen-
namic linker somehow has to be told where the applica- cies; in most (all?) implementations this does not scale
tion can be found and where control has to be transferred linearly.

4 Version 4.1.2 How To Write Shared Libraries

The relocation process is normally3 the most expensive The result of any relocation will be stored somewhere
part of the dynamic linker’s work. It is a process which is in the object with the reference. Ideally and generally
asymptotically at least O(R +nr) where R is the number the target location is in the data segment. If code is in-
of relative relocations, r is the number of named reloca- correctly generated by the user, compiler, or linker relo-
tions, and n is the number of participating DSOs (plus cations might modify text or read-only segments. The
the main executable). Deficiencies in the ELF hash ta- dynamic linker will handle this correctly if the object is
ble function and various ELF extensions modifying the marked, as required by the ELF specification, with the
symbol lookup functionality may well increase the factor DF TEXTREL set in the DT FLAGS entry of the dynamic
to O(R + rn log s) where s is the number of symbols. section (or the existence of the DT TEXTREL flag in old
This should make clear that for improved performance it binaries). But the result is that the modified page can-
is significant to reduce the number if relocations and sym- not be shared with other processes using the same object.
bols as much as possible. After explaining the relocation The modification process itself is also quite slow since
process we will do some estimates for actual numbers. the kernel has to reorganize the memory handling data
structures quite a bit.
1.5.1 The Relocation Process
1.5.2 Symbol Relocations

Relocation in this context means adjusting the application

and the DSOs, which are loaded as the dependencies, to The dynamic linker has to perform a relocation for all
their own and all other load addresses. There are two symbols which are used at run-time and which are not
kinds of dependencies: known at link-time to be defined in the same object as the
reference. Due to the way code is generated on some ar-
chitectures it is possible to delay the processing of some
• Dependencies to locations which are known to be relocations until the references in question are actually
in the own object. These are not associated with a used. This is on many architectures the case for calls
specific symbol since the linker knows the relative to functions. All other kinds of relocations always have
position of the location in the object. to be processed before the object can be used. We will
ignore the lazy relocation processing since this is just a
Note that applications do not have relative reloca- method to delay the work. It eventually has to be done
tions since the load address of the code is known and so we will include it in our cost analysis. To actu-
at link-time and therefore the static linker is able to ally perform all the relocations before using the object is
perform the relocation. used by setting the environment variable LD BIND NOW to
a non-empty value. Lazy relocation can be disabled for
• Dependencies based on symbols. The reference of an individual object by adding the -z now option to the
the definition is generally, but not necessarily, in a linker command line. The linker will set the DF BIND NOW
different object than the definition. flag in the DT FLAGS entry of the dynamic section to
mark the DSO. This setting cannot be undone without
relinking the DSOs or editing the binary, though, so this
The implementation of relative relocations is easy. The
option should only be used if it is really wanted.
linker can compute the offset of the target destination in
the object file at link-time. To this value the dynamic
The actual lookup process is repeated from start for each
linker only has to add the load address of the object and
symbol relocation in each loaded object. Note that there
store the result in the place indicated by the relocation. At
can be many references to the same symbol in differ-
runtime the dynamic linker has to spend only a very small
ent objects. The result of the lookup can be different
and constant amount of time which does not increase if
for each of the objects so there can be no short cuts ex-
more DSOs are used.
cept for caching results for a symbol in each object in
case more than one relocation references the same sym-
The relocation based on a symbol is much more compli-
bol. The lookup scope mentioned in the steps below is an
cated. The ELF symbol resolution process was designed
ordered list of a subset of the loaded objects which can
very powerful so that it can handle many different prob-
be different for each object itself. The way the scope is
lems. All this powerful functionality adds to the com-
computed is quite complex and not really relevant here so
plexity and run-time costs, though. Readers of the fol-
we refer the interested reader to the ELF specification and
lowing description might question the decisions which
section 1.5.4. Important is that the length of the scope is
led to this process. We cannot argue about this here; read-
normally directly dependent on the number of loaded ob-
ers are referred to discussions of ELF. Fact is that symbol
jects. This is another factor where reducing the number
relocation is a costly process and the more DSOs partic-
of loaded objects is increasing performance.
ipate or the more symbols are defined in the DSOs, the
longer the symbol lookup takes.
There are today two different methods for the lookup pro-
3 We ignore the pre-linking support here which in many cases can cess for a symbol. The traditional ELF method proceeds
reduce significantly or even eliminate the relocation costs. in the following steps:

Ulrich Drepper Version 4.1.2 5

Histogram for bucket list length in section [ 2] ’.hash’ (total of 1023 buckets):
Addr: 0x42000114 Offset: 0x000114 Link to section: [ 3] ’.dynsym’
Length Number % of total Coverage
0 132 12.9%
1 310 30.3% 15.3%
2 256 25.0% 40.6%
3 172 16.8% 66.0%
4 92 9.0% 84.2%
5 46 4.5% 95.5%
6 14 1.4% 99.7%
7 1 0.1% 100.0%
Average number of tests: successful lookup: 1.994080
unsuccessful lookup: 1.981427

Figure 3: Example Output for eu-readelf -I libc.so

Histogram for bucket list length in section [ 2] ’.hash’ (total of 191 buckets):
Addr: 0x00000114 Offset: 0x000114 Link to section: [ 3] ’.dynsym’
Length Number % of total Coverage
0 103 53.9%
1 71 37.2% 67.0%
2 16 8.4% 97.2%
3 1 0.5% 100.0%
Average number of tests: successful lookup: 1.179245
unsuccessful lookup: 0.554974

Figure 4: Example Output for eu-readelf -I libnss files.so

1. Determine the hash value for the symbol name. Note that there is no problem if the scope contains more
than one definition of the same symbol. The symbol
2. In the first/next object in the lookup scope: lookup algorithm simply picks up the first definition it
finds. Note that a definition in a DSO being weak has no
2.a Determine the hash bucket for the symbol us-
effects. Weak definitions only play a role in static linking.
ing the hash value and the hash table size in
Having multiple definitions has some perhaps surprising
the object.
consequences. Assume DSO ‘A’ defines and references
2.b Get the name offset of the symbol and using an interface and DSO ‘B’ defines the same interface. If
it as the NUL-terminated name. now ‘B’ precedes ‘A’ in the scope, the reference in ‘A’
2.c Compare the symbol name with the reloca- will be satisfied by the definition in ‘B’. It is said that
tion name. the definition in ‘B’ interposes the definition in ‘A’. This
concept is very powerful since it allows more special-
2.d If the names match, compare the version names ized implementation of an interface to be used without
as well. This only has to happen if both, the replacing the general code. One example for this mech-
reference and the definition, are versioned. It anism is the use of the LD PRELOAD functionality of the
requires a string comparison, too. If the ver- dynamic linker where additional DSOs which were not
sion name matches or no such comparison present at link-time are introduced at run-time. But inter-
is performed, we found the definition we are position can also lead to severe problems in ill-designed
looking for. code. More in this in section 1.5.4.
2.e If the definition does not match, retry with the
next element in the chain for the hash bucket. Looking at the algorithm it can be seen that the perfor-
mance of each lookup depends, among other factors, on
2.f If the chain does not contain any further ele-
the length of the hash chains and the number of objects
ment there is no definition in the current ob-
in the lookup scope. These are the two loops described
ject and we proceed with the next object in
above. The lengths of the hash chains depend on the
the lookup scope.
number of symbols and the choice of the hash table size.
3. If there is no further object in the lookup scope the Since the hash function used in the initial step of the algo-
lookup failed. rithm must never change these are the only two remaining
variables. Many linkers do not put special emphasis on

6 Version 4.1.2 How To Write Shared Libraries

selecting an appropriate table size. The GNU linker tries in a string table associated with the symbol table data
to optimize the hash table size for minimal lengths of the structures. Strings are stored in the C-format; they are
chains if it gets passed the -O option (note: the linker, not terminated by a NUL byte and no initial length field is
the compiler, needs to get this option). used. This means string comparisons has to proceed until
a non-matching character is found or until the end of the
A note on the current implementation of the hash table op-
timization. The GNU binutils linker has a simple minded string. This approach is susceptible to long strings with
heuristic which often favors small table sizes over short common prefixes. Unfortunately this is not uncommon.
chain length. For large projects this might very well in-
crease the startup costs. The overall memory consumption
will be sometimes significantly reduced which might com-
pensate sooner or later but it is still advised to check the namespace some_namespace {
effectiveness of the optimization. A new linker implemen- class some_longer_class_name {
tation is going to be developed and it contains a better algo- int member_variable;
rithm. public:
some_longer_class_name (int p);
int the_getter_function (void);
To measure the effectiveness of the hashing two numbers };
are important: }

• The average chain length for a successful lookup.

• The average chain length for an unsuccessful lookup. The name mangling scheme used by the GNU C++ com-
piler before version 3.0 used a mangling scheme which
put the name of a class member first along with a descrip-
It might be surprising to talk about unsuccessful lookups tion of the parameter list and following it the other parts
here but in fact they are the rule. Note that “unsuccess- of the name such as namespaces and nested class names.
ful” means only unsuccessful in the current objects. Only The result is a name which distinguishable in the begin-
for objects which implement almost everything they get ning if the member names are different. For the example
looked in for is the successful lookup number more im- above the mangled names for the two members functions
portant. In this category there are basically only two ob- look like this figure 5.
jects on a Linux system: the C library and the dynamic
linker itself. In the new mangling scheme used in today’s gcc versions
and all other compilers which are compatible with the
Some versions of the readelf program compute the value common C++ ABI the names start with the namespaces
directly and the output is similar to figures 3 and 4. The and class names and end with the member names. Fig-
data in these examples shows us a number of things. Based ure 6 shows the result for the little example. The mangled
on the number of symbols (2027 versus 106) the chosen names for the two member functions differs only after the
table size is radically different. For the smaller table the 43rd character. This is really bad performance-wise if the
linker can afford to “waste” 53.9% of the hash table en- two symbols should fall into the same hash bucket.4
tries which contain no data. That’s only 412 bytes on
a gABI-compliant system. If the same amount of over- Ada has similar problems. The standard Ada library for
head would be allowed for the libc.so binary the table gcc has all symbols prefixed with ada , then the pack-
would be 4 kilobytes or more larger. That is a big dif- age and sub-package names, followed by function name.
ference. The linker has a fixed cost function integrated Figure 7 shows a short excerpt of the list of symbols from
which takes the table size into account. the library. The first 23 character are the same for all the
names.
The increased relative table size means we have signifi-
cantly shorter hash chains. This is especially true for the The length of the strings in both mangling schemes is
average chain length for an unsuccessful lookup. The av- worrisome since each string has to be compared com-
erage for the small table is only 28% of that of the large pletely when the symbol itself is searched for. The names
table. in the example are not extra ordinarily long either. Look-
ing through the standard C++ library one can find many
What these numbers should show is the effect of reduc- names longer than 120 characters and even this is not the
ing the number of symbols in the dynamic symbol ta- longest. Other system libraries feature names longer than
ble. With significantly fewer symbols the linker has a 200 characters and complicated, “well designed” C++
much better chance to counter the effects of the subopti- projects with many namespaces, templates, and nested
mal hashing function. classes can feature names with more than 1,000 charac-
4 Some people suggested “Why not search from the back?”. Think
Another factor in the cost of the lookup algorithm is con- about it, these are C strings, not PASCAL strings. We do not know the
nected with the strings themselves. Simple string com- length and therefore would have to read every single character of the
parison is used on the symbol names which are stored string to determine the length. The result would be worse.

Ulrich Drepper Version 4.1.2 7

__Q214some_namespace22some_longer_class_namei
the_getter_function__Q214some_namespace22some_longer_class_name

Figure 5: Mangled names using pre-gcc 3 scheme

_ZN14some_namespace22some_longer_class_nameC1Ei
_ZN14some_namespace22some_longer_class_name19the_getter_functionEv

Figure 6: Mangled names using the common C++ ABI scheme

ada__calendar__delays___elabb
ada__calendar__delays__timed_delay_nt
ada__calendar__delays__to_duration

Figure 7: Names from the standard Ada library

ters. One plus point for design, but minus 100 points for looked at it would even be possible to get more accurate
performance. results by multiplying with the exact hash chain length
for the object.
With the knowledge of the hashing function and the de-
tails of the string lookup let us look at a real-world exam- Changing any of the factors ‘number of exported sym-
ple: OpenOffice.org. The package contains 144 separate bols’, ‘length of the symbol strings’, ‘number and length
DSOs. During startup about 20,000 relocations are per- of common prefixes’,‘number of DSOs’, and ‘hash table
formed. Many of the relocations are performed as the size optimization’ can reduce the costs dramatically. In
result of dlopen calls and therefore cannot be optimized general the percentage spent on relocations of the time
away by using prelink [7]. The number of string compar- the dynamic linker uses during startup is around 50-70%
isons needed during the symbol resolution can be used if the binary is already in the file system cache, and about
as a fair value for the startup overhead. We compute an 20-30% if the file has to be loaded from disk.5 It is there-
approximation of this value now. fore worth spending time on these issues and in the re-
mainder of the text we will introduce methods to do just
The average chain length for unsuccessful lookup in all that. So far to remember: pass -O1 to the linker to gener-
DSOs of the OpenOffice.org 1.0 release on IA-32 is 1.1931. ate the final product.
This means for each symbol lookup the dynamic linker
has to perform on average 72 × 1.1931 = 85.9032 string 1.5.3 The GNU-style Hash Table
comparisons. For 20,000 symbols the total is 1,718,064
string comparisons. The average length of an exported
symbol defined in the DSOs of OpenOffice.org is 54.13. All the optimizations proposed in the previous section
Even if we are assuming that only 20% of the string is still leave symbol lookup as a significant factor. A lot
searched before finding a mismatch (which is an opti- of data has to be examined and loading all this data in the
mistic guess since every symbol name is compared com- CPU cache is expensive. As mentioned above, the orig-
pletely at least once to match itself) this would mean a to- inal ELF hash table handling has no more flexibility so
tal of more then 18.5 million characters have to be loaded any solution would have to replace it. This is what the
from memory and compared. No wonder that the startup GNU-style hash table handling does. It can peacefully
is so slow, especially since we ignored other costs. coexist with the old-style hash table handling by having
its own dynamic section entry (DT GNU HASH). Updated
To compute number of lookups the dynamic linker per- dynamic linkers will use the new hash table insted of the
forms one can use the help of the dynamic linker. If the old, therefore provided completely transparent backward
environment variable LD DEBUG is set to symbols one compatibility support. The new hash table implementa-
only has to count the number of lines which start with tion, like the old, is self-contained in each executable and
symbol=. It is best to redirect the dynamic linker’s out- DSO so it is no problem to have binaries with the new
put into a file with LD DEBUG OUTPUT. The number of and some with only the old format in the same process.
string comparisons can then be estimate by multiplying
the count with the average hash chain length. Since the The main cost for the lookup, especially for certain bina-
collected output contains the name of the file which is 5 These numbers assume pre-linking is not used.

8 Version 4.1.2 How To Write Shared Libraries

ries, is the comparison of the symbol names. If the num- This new process seems more complicated. Not only is
ber of comparisons which actually have to be performed this not really the case, it is also much faster. The number
can be reduced we can gain big. A second possible opti- of times we actually have to compare strings is reduced
mization is the layout of the data. The old-style hash table significantly. The Bloom filter alone usually filters out
with its linked list of symbol table entries is not necessar- 80% or more (in many cases 90+%) of all lookups. I.e.,
ily good to the CPU cache. CPU caches work particularly even in the case the hash chains are long no work is done
well when the used memory locations are consecutive. A since the Bloom filter helps to determine that there will be
linked list can jump wildly around and render CPU cache no match. This is done with one signal memory access.
loading and prefetching less effective.
Second, comparing the hash value with that of the symbol
The GNU-style hash tables set out to solve these prob- table entry prevents yet more string comparisons. Each
lem and more. Since compatibility with existing runtime hash chain can contain entries with different hash value
environments could be maintained by providing the old- and this simple word comparison can filter out a lot of
style hash tables in parallel no restrictions of the changes duplicates. There are rarely two entries with the same
were needed. The new lookup process is therefore slightly hash value in a hash chain which means that an unsuc-
different: cessful string comparison is rare. The probability for this
is also increased by using a different hash function than
the original ELF specification dictates. The new function
1. Determine the hash value for the symbol name. is much better in spreading the values out over the value
range of 32-bit values.
2. In the first/next object in the lookup scope:

2.a The hash value is used to determine whether The hash chain array is organized to have all entries for
an entry with the given hash value is present the same hash bucket follow each other. There is no
at all. This is done with a 2-bit Bloom fil- linked list and therefore the cache utilization is much bet-
ter.6 . If the filter indicates there is no such ter.
definition the next object in the lookup scope
is searched. Only if the Bloom filter and the hash function test suc-
ceed do we access the symbol table itself. All symbol ta-
2.b Determine the hash bucket for the symbol us-
ble entries for a hash chain are consecutive, too, so in case
ing the hash value and the hash table size in
we need to access more than one entry the CPU cache
the object. The value is a symbol index.
prefetching will help here, too.
2.c Get the entry from the chain array correspond-
ing to the symbol index. Compare the value One last change over the old format is that the hash ta-
with the hash value of the symbol we are try- ble only contains a few, necessary records for undefined
ing to locate. Ignore bit 0. symbol. The majority of undefined symbols do not have
2.d If the hash value matches, get the name off- to appear in the hash table. This in some cases signif-
set of the symbol and using it as the NUL- icantly reduces the possibility of hash collisions and it
terminated name. certainly increases the Bloom filter success rate and re-
duces the average hash chain length. The result are sig-
2.e Compare the symbol name with the reloca-
nificant speed-ups of 50% or more in code which cannot
tion name.
depend on pre-linking [7] (pre-linking is always faster).
2.f If the names match, compare the version names
as well. This only has to happen if both, the This does not mean, though, that the optimization tech-
reference and the definition, are versioned. It niques described in the previous section are irrelevant.
requires a string comparison, too. If the ver- They still should very much be applied. Using the new
sion name matches or no such comparison hash table implementation just means that not optimizing
is performed, we found the definition we are the exported and referenced symbols will not have a big
looking for. effect on performance as it used to have.
2.g If the definition does not match and the value
loaded from the hash bucket does not have
bit 0 set, continue with the next entry in the The new hash table format was introduced in Fedora
hash bucket array. Core 6. The entire OS, with a few deliberate exceptions,
is created without the compatibility hash table by using
2.h If bit 0 is set there are no further entry in the
--hash-style=gnu. This means the binaries cannot be
hash chain we proceed with the next object in
used on systems without support for the new hash table for-
the lookup scope. mat in the dynamic linker. Since this is never a goal for any
of the OS releases making this decision was a no-brainer.
3. If there is no further object in the lookup scope the
The result is that all binaries are smaller than with the sec-
lookup failed.
ond set of hash tables and in many cases even smaller than
6 http://en.wikipedia.org/wiki/Bloom filter binaries using only the old format.

Ulrich Drepper Version 4.1.2 9

Going back to the OpenOffice.org example, we can make entire global lookup scope is searched before the dynam-
some estimates about the speedup. If the Bloom filter is ically loaded object and its dependencies so that defini-
able to filter out a low 80% of all lookups and the prob- tions would be found first in the global lookup scope ob-
ability of duplicates hash values is a high 15% we only ject before definitions in the local lookup scope. If the
have to actually compare on average 72 × 0.2 × 0.15 × dynamic linker does the lookup as part of a relocation
1.1931 = 2.58 strings. This is an improvement of a fac- this additional dependency is usually taken care of auto-
tor of 33. Adding to this to improved memory handling matically, but this cannot be arranged if the user looks up
and respect for the CPU cache we have even higher gains. symbols in the lookup scope with dlsym.
In real world examples we can reduce the lookup costs so
that programs start up 50% faster or more. And usually there is no reason to use RTLD GLOBAL. For
reasons explained later it is always highly advised to cre-
1.5.4 Lookup Scope ate dependencies with all the DSOs necessary to resolve
all references. RTLD GLOBAL is often used to provide im-
plementations which are not available at link time of a
The lookup scope has so far been described as an ordered DSO. Since this should be avoided the need for this flag
list of most loaded object. While this is correct it has also should be minimal. Even if the programmer has to jump
been intentionally vague. It is now time to explain the through some hoops to work around the issues which are
lookup scope in more detail. solved by RTLD GLOBAL it is worth it. The pain of debug-
ging and working around problems introduced by adding
The lookup scope consists in fact of up to three parts. objects to the global lookup scope is much bigger.
The main part is the global lookup scope. It initially
consists of the executable itself and all its dependencies. The dynamic linker in the GNU C library knows since
The dependencies are added in breadth-first order. That September 2004 one more extension. This extension helps
means first the dependencies of the executable are added to deal with situations where multiple definitions of sym-
in the order of their DT NEEDED entries in the executable’s bols with the same name are not compatible and there-
dynamic section. Then the dependencies of the first de- fore cannot be interposed and expected to work. This is
pendency are added in the same fashion. DSOs already usally a sign of design failures on the side of the peo-
loaded are skipped; they do not appear more than once ple who wrote the DSOs with the conflicting definitions
on the list. The process continues recursively and it will and also failure on the side of the application writer who
stop at some point since there are only a limited number depends on these incompatible DSOs. We assume here
of DSOs available. The exact number of DSOs loaded that an application app is linked with a DSO libone.so
this way can vary widely. Some executables depend on which defines a symbol duplicate and that it dynami-
only two DSOs, others on 200. cally loads a DSO libdynamic.so which depends on
another DSO libtwo.so which also defines a symbol
If an executable has the DF SYMBOLIC flag set (see sec- duplicate. When the application starts it might have a
tion 2.2.7) the object with the reference is added in front global scope like this:
of the global lookup scope. Note, only the object with
the reference itself is added in front, not its dependen-
cies. The effects and reasons for this will be explained app → libone.so → libdl.so → libc.so
later.

A more complicated modification of the lookup scope If now libtwo.so is loaded, the additional local scope
happens when DSOs are loaded dynamic using dlopen. could be like this:
If a DSO is dynamically loaded it brings in its own set
of dependencies which might have to be searched. These libdynamic.so → libtwo.so → libc.so
objects, starting with the one which was requested in the
dlopen call, are appended to the lookup scope if the
object with the reference is among those objects which This local scope is searched after the global scope, pos-
have been loaded by dlopen. That means, those objects sibly with the exception of libdynamic.so which is
are not added to the global lookup scope and they are searched first for lookups in this very same DSO if the
not searched for normal lookups. This third part of the DF DYNAMIC flag is used. But what happens if the sym-
lookup scope, we will call it local lookup scope, is there- bol duplicate is required in libdynamic.so? After
fore dependent on the object which has the reference. all we said so far the result is always: the definition in
libone.so is found since libtwo.so is only in the lo-
The behavior of dlopen can be changed, though. If the cal scope which is searched after the global scope. If the
function gets passed the RTLD GLOBAL flag, the loaded two definitions are incompatible the program is in trou-
object and all the dependencies are added to the global ble.
scope. This is usually a very bad idea. The dynami-
cally added objects can be removed and when this hap- This can be changed with a recent enough GNU C library
pens the lookups of all other objects is influenced. The by ORing RTLD DEEPBIND to the flag word passed as the

10 Version 4.1.2 How To Write Shared Libraries

second parameter to dlopen. If this happens, the dy- extern int bar (int);
namic linker will search the local scope before the global int call_bar (void) {
scope for all objects which have been loaded by the call return bar (foo);
to dlopen. For our example this means the search or- }
der changes for all lookups in the newly loaded DSOs
libdynamic.so and libtwo.so, but not for libc.so
since this DSO has already been loaded. For the two af-
fected DSOs a reference to duplicate will now find the The call to bar requires two relocations: one to load the
definition in libtwo.so. In all other DSOs the definition value of foo and another one to find the address of bar.
in libone.so would be found. If the code would be generated knowing the addresses of
the variable and the function the assembler instructions
While this might sound like a good solution for handling would directly load from or jump to the address. For IA-
compatibility problems this feature should only be used 32 the code would look like this:
if it cannot be avoided. There are several reasons for this:

pushl foo
• The change in the scope affects all symbols and all
call bar
the DSOs which are loaded. Some symbols might
have to be interposed by definitions in the global
scope which now will not happen.
This would encode the addresses of foo and bar as part
• Already loaded DSOs are not affected which could
of the instruction in the text segment. If the address is
cause unconsistent results depending on whether
only known to the dynamic linker the text segment would
the DSO is already loaded (it might be dynamically
have to be modified at run-time. According to what we
loaded, so there is even a race condition).
learned above this must be avoided.
• LD PRELOAD is ineffective for lookups in the dy-
namically loaded objects since the preloaded ob- Therefore the code generated for DSOs, i.e., when using
jects are part of the global scope, having been added -fpic or -fPIC, looks like this:
right after the executable. Therefore they are looked
at only after the local scope.
• Applications might expect that local definitions are movl foo@GOT(%ebx), %eax
always preferred over other definitions. This (and pushl (%eax)
the previous point) is already partly already a prob- call bar@PLT
lem with the use of DF SYMBOLIC but since this
flag should not be used either, the arguments are
still valid. The address of the variable foo is now not part of the in-
• If any of the implicitly loaded DSOs is loaded ex- struction. Instead it is loaded from the GOT. The address
plicitly afterward, its lookup scope will change. of the location in the GOT relative to the PIC register
value (%ebx) is known at link-time. Therefore the text
• Lastly, the flag is not portable. segment does not have to be changed, only the GOT.7

The situation for the function call is similar. The function

The RTLD DEEPBIND flag should really only be used as bar is not called directly. Instead control is transferred
a last resort. Fixing the application to not depend on the to a stub for bar in the PLT (indicated by bar@PLT). For
flag’s functionality is the much better solution. IA-32 the PLT itself does not have to be modified and can
be placed in a read-only segment, each entry is 16 bytes
1.5.5 GOT and PLT in size. Only the GOT is modified and each entry consists
of 4 bytes. The structure of the PLT in an IA-32 DSO
looks like this:
The Global Offset Table (GOT) and Procedure Linkage
Table (PLT) are the two data structures central to the ELF
run-time. We will introduce now the reasons why they
are used and what consequences arise from that. .PLT0:pushl 4(%ebx)
jmp *8(%ebx)
Relocations are created for source constructs like nop; nop
7 There is one more advantage of using this scheme. If the instruc-

tion would be modified we would need one relocation per load/store

instruction. By storing the address in the GOT only one relocation is
extern int foo; needed.

Ulrich Drepper Version 4.1.2 11

 nop; nop • every use of a global variable which is exported
.PLT1:jmp *name1@GOT(%ebx) uses a GOT entry and loads the variable values in-
pushl $offset1 directly;
jmp .PLT0@PC
.PLT2:jmp *name2@GOT(%ebx) • each function which is called (as opposed to refer-
pushl $offset2 enced as a variable) which is not guaranteed to be
jmp .PLT0@PC defined in the calling object requires a PLT entry.
The function call is performed indirectly by trans-
ferring control first to the code in the PLT entry
which in turn calls the function.
This shows three entries, there are as many as needed,
all having the same size. The first entry, labeled with • for some architectures each PLT entry requires at
.PLT0, is special. It is used internally as we will see. least one GOT entry.
All the following entries belong to exactly one function
symbol. The first instruction is an indirect jump where
the address is taken from a slot in the GOT. Each PLT en- Avoiding a jump through the PLT therefore removes on
try has one GOT slot. At startup time the dynamic linker IA-32 16 bytes of text and 4 bytes of data. Avoiding the
fills the GOT slot with the address pointing to the sec- GOT when accessing a global variable saves 4 bytes of
ond instruction of the appropriate PLT entry. I.e., when data and one load instruction (i.e., at least 3 bytes of code
the PLT entry is used for the first time the jump ends at and cycles during the execution). In addition each GOT
the following pushl instruction. The value pushed on entry has a relocation associated with the costs described
the stack is also specific to the PLT slot and it is the off- above.
set of the relocation entry for the function which should
be called. Then control is transferred to the special first 1.5.6 Running the Constructors
PLT entry which pushes some more values on the stack
and finally jumps into the dynamic linker. The dynamic
linker has do make sure that the third GOT slot (offset Once the relocations are performed the DSOs and the ap-
8) contains the address of the entry point in the dynamic plication code can actually be used. But there is one more
linker. Once the dynamic linker has determined the ad- thing to do: optionally the DSOs and the application must
dress of the function it stores the result in the GOT entry be initialized. The author of the code can define for each
which was used in the jmp instruction at the beginning of object a number of initialization functions which are run
the PLT entry before jumping to the found function. This before the DSO is used by other code. To perform the ini-
has the effect that all future uses of the PLT entry will not tialization the functions can use code from the own object
go through the dynamic linker, but will instead directly and all the dependencies. To make this work the dynamic
transfer to the function. The overhead for all but the first linker must make sure the objects are initialized in the
call is therefore “only” one indirect jump. correct order, i.e., the dependencies of an object must be
initialized before the object.
The PLT stub is always used if the function is not guaran-
teed to be defined in the object which references it. Please To guarantee that the dynamic linker has to perform a
note that a simple definition in the object with the refer- topological sort in the list of objects. This sorting is no
ence is not enough to avoid the PLT entry. Looking at linear process. Like all sorting algorithms the run-time is
the symbol lookup process it should be clear that the def- at least O(n log n) and since this is actually a topological
inition could be found in another object (interposition) sort the value is even higher. And what is more: since
in which case the PLT is needed. We will later explain the order at startup need not be the same as the order
exactly when and how to avoid PLT entries. at shutdown (when finalizers have to be run) the whole
process has to be repeated.
How exactly the GOT and PLT is structured is architecture-
specific, specified in the respective psABI. What was said So we have again a cost factor which is directly depend-
here about IA-32 is in some form applicable to some ing on the number of objects involved. Reducing the
other architectures but not for all. For instance, while the number helps a bit even though the actual costs are nor-
PLT on IA-32 is read-only it must be writable for other mally much less than that of the relocation process.
architectures since instead of indirect jumps using GOT
values the PLT entries are modified directly. A reader At this point it is useful to look at the way to correctly
might think that the designers of the IA-32 ABI made a write constructors and destructors for DSOs. Some sys-
mistake by requiring a indirect, and therefore slower, call tems had the convention that exported functions named
instead of a direct call. This is no mistake, though. Hav- init and fini are automatically picked as constructor
ing a writable and executable segment is a huge security and destructor respectively. This convention is still fol-
problem since attackers can simply write arbitrary code lowed by GNU ld and using functions with these names
into the PLT and take over the program. We can anyhow on a Linux system will indeed cause the functions used
summarize the costs of using GOT and PLT like this: in these capacities. But this is totally, 100% wrong!

12 Version 4.1.2 How To Write Shared Libraries

By using these functions the programmer overwrites what- • Related, the application and the dependencies
ever initialization and destruction functionality the sys- with additional dependencies must record the
tem itself is using. The result is a DSO which is not names of the dependencies. This is not a ter-
fully initialized and this sooner or later leads to a catas- ribly high cost but certainly can sum up if
trophy. The correct way of adding constructors and de- there are many dozens of dependencies.
structors is by marking functions with the constructor • The lookup scope grows. This is one of the
and destructor function attribute respectively. dominating factors in cost equation for the re-
locations.
• More objects means more symbol tables which
void in turn normally means more duplication. Un-
__attribute ((constructor)) defined references are not collapsed into one
init_function (void) and handling of multiple definitions have to
{ be sorted out by the dynamic linker.
... Moreover, symbols are often exported from a
}
DSO to be used in another one. This would
not have to happen if the DSOs would be merged.
void
__attribute ((destructor)) • The sorting of initializers/finalizers is more
fini_function (void) complicated.
{
• In general does the dynamic linker have some
...
} overhead for each loaded DSO per process.
Every time a new DSO is requested the list of
already loaded DSOs must be searched which
can be quite time consuming since DSOs can
These functions should not be exported either (see sec- have many aliases.
tions 2.2.2 and 2.2.3) but this is just an optimization.
With the functions defined like this the runtime will ar- Number of Symbols The number of exported and unde-
range that they are called at the right time, after perform- fined symbols determines the size of the dynamic
ing whatever initialization is necessary before. symbol table, the hash table, and the average hash
table chain length. The normal symbol table is not
1.6 Summary of the Costs of ELF used at run-time and it is therefore not necessary
to strip a binary of it. It has no impact on perfor-
mance.
We have now discussed the startup process and how it is Additionally, fewer exported symbols means fewer
affected by the form of the binaries. We will now summa- chances for conflicts when using pre-linking (not
rize the various factors so that we later on can determine covered further).
the benefits of an optimization more easily.
Length of Symbol Strings Long symbol lengths cause
often unnecessary costs. A successful lookup of a
symbol must match the whole string and compar-
Code Size As everywhere, a reduced size for code with
ing dozens or hundreds of characters takes time.
the same semantics often means higher efficiency
Unsuccessful lookups suffer if common prefixes
and performance. Smaller ELF binaries need less
are long as in the new C++ mangling scheme. In
memory at run-time.
any case do long symbol names cause large string
In general the compiler will always generate the tables which must be present at run-time and thereby
best code possible and we do not cover this further. is adding costs in load time and in use of address
But it must be known that every DSO includes a space which is an issue for 32-bit machines.
certain overhead in data and code. Therefore fewer
DSOs means smaller text. Number of Relocations Processing relocations constitute
the majority of work during start and therefore any
Number of Objects The fact that a smaller number of reduction is directly noticeable.
objects containing the same functionality is bene-
Kind of Relocations The kind of relocations which are
ficial has been mentioned in several places:
needed is important, too, since processing a rela-
tive relocation is much less expensive than a nor-
• Fewer objects are loaded at run-time. This mal relocation. Also, relocations against text seg-
directly translates to fewer system call. In the ments must be avoided.
GNU dynamic linker implementation loading
a DSO requires at least 8 system calls, all of Placement of Code and Data All executable code should
them can be potentially quite expensive. be placed in read-only memory and the compiler

Ulrich Drepper Version 4.1.2 13

$ env LD_DEBUG=statistics /bin/echo ’+++ some text +++’
...:
...: run-time linker statistics:
...: total startup time in dynamic loader: 748696 clock cycles
...: time needed for relocation: 378004 clock cycles (50.4%)
...: number of relocations: 133
...: number of relocations from cache: 5
...: time needed to load objects: 193372 clock cycles (25.8%)
+++ some text +++
...:
...: run-time linker statistics:
...: final number of relocations: 188
...: final number of relocations from cache: 5

Figure 8: Gather Startup Statistics

normally makes sure this is done correctly. When in the kernel to load/map binaries. In this example the
creating data objects it is mostly up to the user relocation processing dominates the startup costs with
to make sure it is placed in the correct segment. more than 50%. There is a lot of potential for optimiza-
Ideally data is also read-only but this works only tions here. The unit used to measure the time is CPU
for constants. The second best choice is a zero- cycles. This means that the values cannot even be com-
initialized variable which does not have to be ini- pared across different implementations of the same ar-
tialized from file content. The rest has to go into chitecture. E.g., the measurement for a PentiumRM III and
the data segment. a PentiumRM 4 machine will be quite different. But the
measurements are perfectly suitable to measure improve-
ments on one machine which is what we are interested
In the following we will not cover the first two points here.
given here. It is up to the developer of the DSO to de-
cide about this. There are no small additional changes to Since relocations play such a vital part of the startup per-
make the DSO behave better, these are fundamental de- formance some information on the number of relocations
sign decisions. We have voiced an opinion here, whether is printed. In the example a total of 133 relocations are
it is has any effect remains to be seen. performed, from the dynamic linker, the C library, and the
executable itself. Of these 5 relocations could be served
1.7 Measuring ld.so Performance from the relocation cache. This is an optimization imple-
mented in the dynamic linker to handle the case of mul-
tiple relocations against the same symbol more efficient.
To perform the optimizations it is useful to quantify the
After the program itself terminated the same information
effect of the optimizations. Fortunately it is very easy to
is printed again. The total number of relocations here is
do this with glibc’s dynamic linker. Using the LD DEBUG
higher since the execution of the application code caused
environment variable it can be instructed to dump in-
a number, 55 to be exact, of run-time relocations to be
formation related to the startup performance. Figure 8
performed.
shows an example invocation, of the echo program in
this case.
The number of relocations which are processed is stable
across successive runs of the program. The time mea-
The output of the dynamic linker is divided in two parts.
surements not. Even in a single-user mode with no other
The part before the program’s output is printed right be-
programs running there would be differences since the
fore the dynamic linker turns over control to the appli-
cache and main memory has to be accessed. It is there-
cation after having performed all the work we described
fore necessary to average the run-time over multiple runs.
in this section. The second part, a summary, is printed
after the application terminated (normally). The actual
It is obviously also possible to count the relocations with-
format might vary for different architectures. It includes
out running the program. Running readelf -d on the
the timing information only on architectures which pro-
binary shows the dynamic section in which the DT RELSZ,
vide easy access to a CPU cycle counter register (modern
DT RELENT, DT RELCOUNT, and DT PLTRELSZ entries are
IA-32, IA-64, x86-64, Alpha in the moment). For other
interesting. They allow computing the number of normal
architectures these lines are simply missing.
and relative relocations as well as the number of PLT en-
tries. If one does not want to do this by hand the relinfo
The timing information provides absolute values for the
script in appendix A can be used.
total time spend during startup in the dynamic linker, the
time needed to perform relocations, and the time spend

14 Version 4.1.2 How To Write Shared Libraries

2 Optimizations for DSOs the generation of position-independent code.

Which of the two options, -fpic or -fPIC, have to be

In this section we describe various optimizations based used must be decided on a case-by-case basis. For some
on C or C++ variables or functions. The choice of vari- architectures there is no difference at all and people tend
able or function, unless explicitly said, is made deliber- to be careless about the use. For most RISC there is a big
ately since many of the implementations apply to the one difference. As an example, this is the code gcc generates
or the other. But there are some architectures where func- for SPARC to read a global variable global when using
tions are handled like variables. This is mainly the case -fpic:
for embedded RISC architectures like SH-3 and SH-4
which have limitations in the addressing modes they pro-
vide which make it impossible to implement the function
handling as for other architectures. In most cases it is sethi %hi(_GLOBAL_OFFSET_TABLE_-4),%l7
no problem to apply the optimizations for variables and call .LLGETPC0
functions at the same time. This is what in fact should be add %l7,%lo(_GLOBAL_OFFSET_TABLE_+4),%l7
ld [%l7+global],%g1
done all the time to achieve best performance across all
ld [%g1],%g1
architectures.

The most important recommendation is to always use

-fpic or -fPIC when generating code which ends up in And this is the code sequence if -fPIC is used:
DSOs. This applies to data as well as code. Code which
is not compiled this way almost certainly will contain text
relocations. For these there is no excuse. Text relocations
requires extra work to apply in the dynamic linker. And sethi %hi(_GLOBAL_OFFSET_TABLE_-4),%l7
argumentation saying that the code is not shared because call .LLGETPC0
no other process uses the DSO is invalid. In this case it is add %l7,%lo(_GLOBAL_OFFSET_TABLE_+4),%l7
sethi %hi(global),%g1
not useful to use a DSO in the first place; the code should
or %g1,%lo(global),%g1
just be added to the application code. ld [%l7+%g1],%g1
ld [%g1],%g1
Some people try to argue that the use of -fpic/-fPIC
on some architectures has too many disadvantages. This
is mainly brought forward in argumentations about IA-
32. Here the use of %ebx as the PIC register deprives In both cases %l7 is loaded with the address of the GOT
the compiler of one of the precious registers it could use first. Then the GOT is accessed to get the address of
for optimization. But this is really not that much of a global. While in the -fpic case one instruction is suf-
problem. First, not having %ebx available was never a ficient, three instructions are needed in the -fPIC case.
big penalty. Second, in modern compilers (e.g., gcc after The -fpic option tells the compiler that the size of the
release 3.1) the handling of the PIC register is much more GOT does not exceed an architecture-specific value (8kB
flexible. It is not always necessary to use %ebx which in case of SPARC). If only that many GOT entries can
can help eliminating unnecessary copy operations. And be present the offset from the base of the GOT can be
third, by providing the compiler with more information as encoded in the instruction itself, i.e., in the ld instruc-
explained later in this section a lot of the overhead in PIC tion of the first code sequence above. If -fPIC is used
can be removed. This all combined will lead to overhead no such limit exists and so the compiler has to be pes-
which is in most situations not noticeable. simistic and generate code which can deal with offsets of
any size. The difference in the number of instructions in
When gcc is used, the options -fpic/-fPIC also tell the this example correctly suggests that the -fpic should be
compiler that a number of optimizations which are pos- used at all times unless it is absolutely necessary to use
sible for the executable cannot be performed. This has to -fPIC. The linker will fail and write out a message when
do with symbol lookups and cutting it short. Since the this point is reached and one only has to recompile the
compiler can assume the executable to be the first object code.
in the lookup scope it knows that all references of global
symbols known to be defined in the executable are re- When writing assembler code by hand it is easy to miss
solved locally. Access to locally defined variable could cases where position independent code sequences must
be done directly, without using indirect access through be used. The non-PIC sequences look and actually are
the GOT. This is not true for DSOs: the DSOs can be simpler and more natural. Therefore it is extremely im-
later in the lookup scope and earlier objects might be in- portant to in these case to check whether the DSO is
terposed. It is therefore mandatory to compile all code marked to contain text relocations. This is easy enough
which can potentially end up in a DSO with -fpic/-fPIC to do:
since otherwise the DSO might not work correctly. There
is no compiler option to separate this optimization from readelf -d binary | grep TEXTREL

Ulrich Drepper Version 4.1.2 15

If this produces any output text relocations are present tions with zero are treated special by some compil-
and one better starts looking what causes them. ers.

2.1 Data Definitions

Normally there is not much the user has to do to create
optimal ELF files. The compiler will take care of avoid-
Variables can be defined in C and C++ in several different ing the initializers. To achieve the best results even with
ways. Basically there are three kinds of definitions: old compilers it is desirable to avoid explicit initializa-
tions with zero if possible. This creates normally com-
mon variables but if combined with gcc’s -fno-common
Common Common variables are more widely used FOR- flag the same reports about multiple definitions one would
TRAN but they got used in C and C++ as well to get for initialized variables can be seen.
work around mistakes of programmers. Since in
the early days people used to drop the extern key- There is one thing the programmer is responsible for. As
word from variable definitions, in the same way it an example look at the following code:
is possible to drop it from function declaration, the
compiler often has multiple definitions of the same
variable in different files. To help the poor and
clueless programmer the C/C++ compiler normally bool is_empty = true;
generates common variables for uninitialized defi- char s[10];
nitions such as
const char *get_s (void) {
int foo; return is_empty ? NULL : s;
For common variables there can be more than one }
definition and they all get unified into one location
in the output file. Common variables are always
initialized with zero. This means their value does
not have to be stored in an ELF file. Instead the The function get s uses the boolean variable is empty
file size of a segment is chosen smaller than the to decide what to do. If the variable has its initial value
memory size as described in 1.4. the variable s is not used. The initialization value of
is empty is stored in the file since the initialize is non-
Uninitialized If the programmer uses the compiler com- zero. But the semantics of is empty is chosen arbitrar-
mand line option -fno-common the generated code ily. There is no requirement for that. The code could
will contain uninitialized variables instead of com- instead be rewritten as:
mon variables if a variable definition has no ini-
tializer. Alternatively, individual variables can be
marked like this:
bool not_empty = false;
int foo attribute ((nocommon));
char s[10];
The result at run-time is the same as for common
variable, no value is stored in the file. But the rep- const char *get_s (void) {
resentation in the object file is different and it al- return not_empty ? s : NULL;
lows the linker to find multiple definitions and flag }
them as errors. Another difference is that it is pos-
sible to define aliases, i.e., alternative names, for
uninitialized variables while this is not possible for
common variables. Now the semantics of the control variable is reversed. It
is initialized with false which is guaranteed to have the
With recent gcc versions there is another method to numeric value zero. The test in the function get s has
create uninitialized variables. Variables initialized to be changed as well but the resulting code is not less or
with zero are stored this way. Earlier gcc versions more efficient than the old code.
stored them as initialized variables which took up
space in the file. This is a bit cumbersome for vari- By simple transformations like that it is often possible
ables with structured types. So, sticking with the to avoid creating initialized variables and instead using
per-variable attribute is probably the best way. common or uninitialized variables. This saves disk space
Initialized The variable is defined and initialized to a and eventually improves startup times. The transforma-
programmer-defined value. In C: tion is not limited to boolean values. It is sometimes pos-
sible to do it for variables which can take on more than
int foo = 42; two values, especially enumeration values. When defin-
In this case the initialization value is stored in the ing enums one should always put the value, which is most
file. As described in the previous case initializa- often used as initializer, first in the enum definition. I.e.

16 Version 4.1.2 How To Write Shared Libraries

 enum { val1, val2, val3 }; In the discussions of the various methods we will use one
example:
should be rewritten as

enum { val3, val1, val2 };

int last;
if val3 is the value most often used for initializations.
To summarize, it is always preferable to add variables int next (void) {
return ++last;
as uninitialized or initialized with zero as opposed to as
}
initialized with a value other than zero.
int index (int scale) {
2.2 Export Control return next () << scale;
}

When creating a DSO from a collection of object files the

dynamic symbol table will by default contain all the sym-
bols which are globally visible in the object files. In most Compiled on a IA-32 Linux machine a DSO with only
cases this set is far too large. Only the symbols which are this code (plus startup code etc) contains seven reloca-
actually part of the ABI should be exported. Failing to tions, two of which are relative, and four PLT entries (use
restrict the set of exported symbols are numerous draw- the relinfo script). We will see how we can improve on
backs: this. Four of the normal and both relative relocations as
well as three PLT entries are introduced by the additional
code used by the linker to create the DSO. The actual ex-
• Users of the DSO could use interfaces which they ample code creates only one normal relocation for last
are not supposed to. This is problematic in revi- and one PLT entry for next. To increment and read the
sions of the DSO which are meant to be binary variable last in next the compiler generates code like
compatible. The correct assumption of the DSO
developer is that interfaces, which are not part of
the ABI, can be changed arbitrarily. But there are
always users who claim to know better or do not movl last@GOT(%ebx), %edx
care about rules. movl (%edx), %eax
incl %eax
• According to the ELF lookup rules all symbols in movl %eax, (%edx)
the dynamic symbol table can be interposed (un-
less the visibility of the symbol is restricted). I.e.,
definitions from other objects can be used. This
means that local references cannot be bound at link and the call of next is compiled to
time. If it is known or intended that the local defi-
nition should always be used the symbol in the ref-
erence must not be exported or the visibility must
call next@PLT
be restricted.

• The dynamic symbol table and its string table are

available at run-time and therefore must be loaded.
This can require a significant amount of memory, These code fragments were explained in section 1.5.5.
even though it is read-only. One might think that
the size is not much of an issue but if one exam- 2.2.1 Use static
ines the length of the mangled names of C++ vari-
ables or functions, it becomes obvious that this is
not the case. In addition we have the run-time costs The easiest way to not export a variable or function is to
of larger symbol tables which we discussed in the define it with file-local scope. In C and C++ this is done
previous section. by defining it with static, in C++ additionally using
anonymous namespaces. This is for many people obvi-
ous but unfortunately not for all. Many consider adding
We will now present a number of possible solutions for static as optional. This is true when considering only
the problem of exported interfaces. Some of them solve the C semantics of the code.
the same problem in slightly different ways. We will say
which method should be preferred. The programmer has If in our example neither last or next are needed out-
to make sure that whatever is used is available on the tar- side the file we can change the source to:
get system.

Ulrich Drepper Version 4.1.2 17

 file as the references and then define the referenced ob-
jects as static. When generating the production bina-
static int last; ries it might even be desirable to merge as many input
files as possible together to mark as many objects as pos-
static int next (void) { sible static. Unless one is comfortable with one giant
return ++last; file there is a limit on how many functions can be grouped
} together. It is not necessary to continue the process ad in-
finitum since there are other ways to achieve the same
int index (int scale) {
result (minus inlining).
return next () << scale;
}

2.2.2 Define Global Visibility

Compiled in the same way as before we see that all the re- The next best thing to using static is to explicitly de-
locations introduced by our example code vanished. I.e., fine the visibility of objects in the DSO. The generic ELF
we are left with six relocations and three PLT entries. The ABI defines visibility of symbols. The specification de-
code to access last now looks like this: fines four classes of which here only two are of interest.
STV DEFAULT denotes the normal visibility. The symbol
is exported and can be interposed. The other interesting
class is denoted by STV HIDDEN. Symbols marked like
movl last@GOTOFF(%ebx), %eax
incl %eax
this are not exported from the DSO and therefore can-
movl %eax, last@GOTOFF(%ebx) not be used from other objects. There are a number of
different methods to define visibility.

Starting with version 4.0, gcc knows about a the com-

The code improved by avoiding the step which loads the mand line option -fvisibility. It takes a parameter
address of the variable from the GOT. Instead, both mem- and the valid forms are these:
ory accesses directly address the variable in memory. At
link-time the variable location has a fixed offset from the
PIC register, indicated symbolically by last@GOTOFF.
By adding the value to the PIC register value we get the -fvisibility=default
address of last. Since the value is known at link-time -fvisibility=hidden
this construct does not need a relocation at run-time. -fvisibility=internal
-fvisibility=protected
The situation is similar for the call to next. The IA-32
architecture, like many others, knows a PC-relative ad-
dressing mode for jumps and calls. Therefore the com-
piler can generate a simple jump instruction Only the first two should ever be used. The default is
unsurprisingly default since this is the behavior of the
compiler before the introduction of this option. When
-fvisibility=hidden is specified gcc changes the de-
call next fault visibility of all defined symbols which have no ex-
plicit assignment of visibility: all symbols are defined
with STV HIDDEN unless specified otherwise. This op-
tion has to be used with caution since unless the DSO
and the assembler generates a PC-relative call. The dif- is prepared by having all APIs marked as having default
ference between the address of the instruction following visibility, the generated DSO will not have a single ex-
the call and the address of next is constant at link-time ported symbol. This is usually not what is wanted.
and therefore also does not need any relocation. Another
advantage is that, in the case of IA-32, the PIC register In general it is the preference of the author which decides
does not have to be set up before the jump. If the com- whether -fvisibility=hidden should be used. If it
piler wouldn’t know the jump target is in the same DSO is not used, symbols which are not to be exported need
the PIC register would have to be set up. Other architec- to be marked in one way or another. The next section
tures have similar requirements. will go into details. In case the option is used all ex-
ported functions need to be declared as having visibility
The generated code is optimal. The compiler might even default which usually means the header files are sig-
consider inlining some code if it finds that this is bene- nificantly uglified. On the other hand it means that no
ficial. It is always advised that the programmer places symbol can accidentally be exported because an appro-
the various variable and function definitions in the same priate declaration etc is missing. In some situations this

18 Version 4.1.2 How To Write Shared Libraries

can prevent bad surprises.8 next (void) {
return ++last;
2.2.3 Define Per-Symbol Visibility }
#pragma GCC visibility pop

Instead of changing the default visibility the programmer int index (int scale) {
can choose to define to hide individual symbols. Or, if return next () << scale;
the default visibility is hidden, make specific symbols ex- }
portable by setting the visibility to default.

Since the C language does not provide mechanisms to

define the visibility of a function or variable gcc resorts As in the example using the attributes, last and next
once more to using attributes: are both defined with hidden visibility while index is
defined with default visibility (assuming this is the de-
fault currently in use). As the pragma syntax suggests, it
is possible to nest the pragmas with the expected result.
int last
__attribute__ ((visibility ("hidden"))); In case the -fvisibility=hidden command line op-
tion is used, individual symbols can be marked as ex-
int portable by using the same syntax as presented in this
__attribute__ ((visibility ("hidden")))
section, except with default in place of hidden. In
next (void) {
return ++last;
fact, the names of all four visibilities are allowed in the
} attribute or pragma.

int index (int scale) { Beside telling the backend of the compiler to emit code to
return next () << scale; flag the symbol as hidden, changing the visibility has an-
} other purpose: it allows the compiler to assume the defi-
nition is local. This means the addressing of variables and
function can happen as if the definitions would be locally
defined in the file as static. Therefore the same code
This defines the variable last and the function next sequences we have seen in the previous section can be
as hidden. All the object files which make up the DSO generated. Using the hidden visibility attribute is there-
which contains this definition can use these symbols. I.e., fore almost completely equivalent to using static; the
while static restricts the visibility of a symbol to the only difference is that the compiler cannot automatically
file it is defined in, the hidden attribute limits the visibil- inline the function since it need not see the definition.
ity to the DSO the definition ends up in. In the example
above the definitions are marked. This does not cause any We can now refine the rule for using static: merge
harm but it is in any case necessary to mark the declara- source files and mark as many functions static as far as
tion. In fact it is more important that the declarations are one feels comfortable. In any case merge the files which
marked appropriately since it is mainly the code gener- contain functions which potentially can be inlined. In all
ated for in a reference that is influenced by the attribute. other cases mark functions (the declarations) which are
not to be exported from the DSO as hidden.
Instead of adding an visibility attribute to each declara-
tion or definition, it is possible to change the default tem- Note that the linker will not add hidden symbols to the
porarily for all definitions and declarations the compiler dynamic symbol table. I.e., even though the symbol ta-
sees at this time. This is mainly useful in header files bles of the object files contain hidden symbols they will
since it reduces changes to a minimum but can also be disappear automatically. By maximizing the number of
useful for definitions. This compiler feature was also in- hidden declarations we therefore reduce the size of the
troduced in gcc 4.0 and is implemented using a pragma:9 symbol table to the minimum.

The generic ELF ABI defines another visibility mode:

protected. In this scheme references to symbols defined
#pragma GCC visibility push(hidden)
int last;
in the same object are always satisfied locally. But the
symbols are still available outside the DSO. This sounds
int like an ideal mechanism to optimize DSO by avoiding the
use of exported symbols (see section 2.2.7) but it isn’t.
8 Accidentally exporting symbol can mean that programs can use
Processing references to protected symbols is even more
and get dependent on them. Then it is hard to remove the symbol again
for binary compatibility reasons.
expensive than normal lookup. The problem is a require-
9 Note: ISO C99 introduced Pragma which allows using pragmas ment in the ISO C standard. The standard requires that
in macros. function pointers, pointing to the same function, can be

Ulrich Drepper Version 4.1.2 19

compared for equality. This rule would be violated with a member function val are defined as hidden. The sym-
fast and simple-minded implementation of the protected bols cannot be accessed outside the DSO the definitions
visibility. Assume an application which references a pro- appear in. Please note that this is an additional restriction
tected function in a DSO. Also in the DSO is another on top of the C++ access rules. For the member functions
function which references said function. The pointer in one way around the problem is to instantiate the class in
the application points to the PLT entry for the function more than one DSO. This is usually causing no problems
in the application’s PLT. If a protected symbol lookup and “only” adds to code bloat.
would simply return the address of the function inside
the DSO the addresses would differ. Things are getting more interesting when static data mem-
bers or static local variables in member functions are used.
In programming environments without this requirement In this case there must be exactly one definition used
on function pointers the use of the protected visibility (please note: “used”, not “present”). To obey this rule
would be useful and fast. But since there usually is only it is either necessary to not restrict the export of the static
one implementation of the dynamic linker on the sys- data member of member function from the DSO or to
tem and this implementation has to handle C programs make sure all accesses of the data or function are made
as well, the use of protected is highly discouraged. in the DSO with the definitions. If multiple definitions
are present it is very easy to make mistakes when hiding
There are some exceptions to these rules. It is possible static data members or the member functions with static
to create ELF binaries with non-standard lookup scopes. variables since the generated code has no way of knowing
The simplest example is the use of DF SYMBOLIC (or of that there are multiple definitions of the variables. This
DT SYMBOLIC in old-style ELF binaries, see page 25). leads to very hard to debug bugs.
In these cases the programmer decided to create a non-
standard binary and therefore accepts the fact that the In the example code above the static data member u is de-
rules of the ISO C standard do not apply. clared hidden. All users of the member must be defined
in the same DSO. C++ access rules restrict access only to
2.2.4 Define Visibility for C++ Classes member functions, regardless of where they are defined.
To make sure all users are defined in the DSO with the
definition of u it is usually necessary to avoid inline func-
For C++ code we can use the attributes as well but they tions which access the hidden data since the inline gener-
have to be used very carefully. Normal function or vari- ated code can be placed in any DSO which contains code
able definitions can be handled as in C. The extra name using the class definition. The member function offset
mangling performed has no influence on the visibility. is a prime example of a function which should be inlined
The story is different when it comes to classes. The sym- but since it accesses u it cannot be done. Instead offset
bols and code created for class definitions are member is exported as an interface from the DSO which contains
functions and static data or function members. These the definition of u.
variables and functions can easily be declared as hidden
but one has to be careful. First an example of the syntax. If a member function is marked as hidden, as val is in
the example, it cannot be called from outside the DSO.
Note that in the example the compiler allows global ac-
class foo {
cess to the member function since it is defined as a public
static int u __attribute__ member. The linker, not the compiler, will complain if
((visibility ("hidden"))); this member function is used outside the DSO with the
int a; instantiation. Inexperienced or not fully informed users
public: might interpret this problem as a lack of instantiation
foo (int b = 1); which then leads to problems due to multiple definitions.
void offset (int n);
int val () const __attribute__ Because these problems are so hard to debug it is essen-
((visibility ("hidden"))); tial to get the compiler involved in making sure the user
}; follows the necessary rules. The C++ type system is rich
enough to help if the implementor puts some additional
int foo::u __attribute__
((visibility ("hidden")));
effort in it. The key is to mimic the actual symbol access
foo::foo (int b) : a (b) { } as closely as possible with the class definition. For this
void foo::offset (int n) { u = n; } reason the class definitions of the example above should
int actually look like this:
__attribute__ ((visibility ("hidden")))
foo::val () const { return a + u; }

class foo {
static int u __attribute__
In this example code the static data member u and the ((visibility ("hidden")));

20 Version 4.1.2 How To Write Shared Libraries

 int a; a lot of work. Since version 4.0 gcc knows about the op-
public: tion -fvisibility-inlines-hidden which does just
foo (int b = 1); what is wanted. If this option is used a referenced in-
int val () const __attribute__ line function is assumed to be hidden and an out-of-line
((visibility ("hidden"))); copy of the function is marked with STV HIDDEN. I.e., if
void offset (int n); the function is not inlined the separate function created is
}; not exported. This is a quite frequent situation at times
since not all functions the programmer thinks should be
class foo_ext : protected foo {
inlined are eligible according to the compiler’s analysis.
public:
foo_ext (int b = 1) : foo (b) { } This option is usable in almost all situations. Only if the
void offset (int n) functions in the different DSOs can be different or if the
{ return foo::offset (n); } code depends on exactly one copy of the function ever
}; being used (e.g., if the function address is expected to be
the same) should this option be avoided.

If a C++ class is used only for the implementation and

The class foo is regarded as a private class, not to be not used in any interface of a DSO using the code, then
used outside the DSO with the instantiation. The public it would be possible to mark each member function and
interface would be the class foo ext. It provides access static data element as hidden. This is cumbersome, error-
to the two public interfaces of the underlying class. As prone, and incomplete, though. There is perhaps a large
long as the users of the DSO containing the definitions number of members which need to be marked and when a
respect the requirement that only foo ext can be used new member is added it is easy to forget about adding the
there is no way for the compiler not noticing accesses to necessary attributes. The incompleteness stems from the
foo::u and foo::val outside the DSO containing the fact that the C++ compiler automatically generates a few
definitions. members functions such are constructors and destructors.
These member functions would not be affected by the
Template class and functions are not different. The syn- attributes.
tax is the same. Non-inline function definitions get yet
again less readable but that is something which can be The solution to these problems is to explicitly determine
mostly hidden with a few macros. the visibility of the entire class. Since version 4.0 does
gcc have support for this. There are two ways to achieve
the goal. First, the already mentioned pragma can be
used.
template<class T>
class a {
T u;
public: #pragma GCC visibility push(hidden)
a (T a = 0); class foo {
T r () const __attribute__ ...
((visibility ("hidden"))); };
}; #pragma GCC visibility pop

template<class T> a<T>::a (T a)

{ u = a; }
template<class T> T All member functions and static data members of foo
__attribute__ ((visibility ("hidden"))) are automatically defined as hidden. This extends even to
a<T>::r () const { return u; } implicitly generated functions and operators if necessary.

The second possibility is to use yet another extension in

gcc 4.0. It is possible to mark a function as hidden when
For templatized classes the problems of making sure that it is defined. The syntax is this:
if necessary only one definition is used is even harder to
fix due to the various approaches to instantiation.

One sort of function which can safely be kept local and class __attribute ((visibility ("hidden")))
foo {
not exported are inline function, either defined in the class
...
definition or separately. Each compilation unit must have };
its own set of all the used inline functions. And all the
functions from all the DSOs and the executable better be
the same and are therefore interchangeable. It is possible
to mark all inline functions explicitly as hidden but this is Just as with the pragma, all defined functions are defined

Ulrich Drepper Version 4.1.2 21

as hidden symbols. Explicitly using attributes should be ery symbol can belong to one of two classes: exported or
preferred since the effect of the pragmas is not always not exported. Symbols can be listed individually, glob-
obvious. If the push and pop lines are far enough from expressions can be used, or the special * catch-all glob
each other a programmer might accidentally add a new expression can be used. The latter only once. The sym-
declaration in the range even though the visibility of this bol map file for our example code could look like this:
new declaration is not meant to be affected. Both, the
pragma and the class attribute, should only be used in
internal headers. In the headers which are used to ex-
{
pose the API of the DSO it makes no sense to have them global: index;
since the whole point is to hide the implementation de- local: *;
tails. This means it is always a good idea to differentiate };
between internal and external header files.

Defining entire classes with hidden visibility has some

problems which cannot be modeled with sophisticated This tells the linker that the symbol index is to be ex-
class layout or moving the definition in private headers. ported and all others (matched by *) are local. We could
For exception handling the compiler generates data struc- have listed last and next explicitly in the local: list
tures (typeinfo symbols) which are also marked ac- but it is generally advised to always use the catch-all case
cording to the visibility attribute used. If an object of to mark all not explicitly mentioned symbols as local.
this type is thrown the catch operation has to look for This avoids surprises by allowing access only to the sym-
the typeinfo information. If that information is in a bols explicitly mentions. Otherwise there would also be
different DSO the search will be unsuccessful and the the problem with symbols which are matched neither by
program will terminate. All classes which are used in the global: nor by the local:, resulting in unspecified
exception handling and where the throw and catch are behavior. Another unspecified behavior is if a name ap-
not both guaranteed to reside in the DSO with the defi- pears in both lists or is matched using globbing in both
nition must be declared with default visibility. Individual lists.
members can still be marked with an visibility attribute
but since the typeinfo data is synthesized by the com- To generate a DSO with this method the user has to pass
piler on command there is no way for the programmer to the name of the map file with the --version-script
overwrite a hidden visibility attribute associated with the option of the linker. The name of the option suggests that
class. the scripts can be used for more. We will get back to this
when we discuss ABIs in the next chapter.
The use of the most restrictive visibility possible can be
of big benefit for C++ code. Each inline function which
is (also) available as a stand-alone function, every syn- $ gcc -shared -o foo.so foo.c -fPIC \
thesized function are variable has a symbol associated -Wl,--version-script=foo.map
which is by default exported. For templatized classes this
is even worse, since each instantiated class can bring is
many more symbols. It is best to design the code right
away so that the visibility attributes can be applied when- The file foo.map is supposed to contain text like the one
ever possible. Compatibility with older compilers can shown above.
easily be achieved by using macros.
It is of course also possible to use export maps with C++
2.2.5 Use Export Maps code. One has two options in this case: explicitly name
the symbols using their mangled names, or rely on pat-
tern matching for the mangled names. Using the man-
If for one reason or another none of the previous two so- gled names is straight-forwarded. Just use the identifiers
lutions are applicable the next best possibility is to in- as in the C examples. Using the demangled names re-
struct the linker to do something. Only the GNU and quire support in the linker. Assume a file defining the
Solaris linker are known to support this, at least with the following functions:
syntax presented here. Using export maps is not only
useful for the purpose discussed here. When discussing
maintenance of APIs and ABIs in chapter 3 the same kind
int foo (int a) { ... }
of input file is used. This does not mean the previous two int bar (int a) { ... }
methods should not be preferred. Instead, export (and struct baz {
symbol) maps can and should always be used in addition baz (int);
to the other methods described. int r () const;
int s (int);
The concept of export maps is to tell the linker explicitly };
which symbols to export from the generated object. Ev-

22 Version 4.1.2 How To Write Shared Libraries

A DSO containing definitions for all these functions and DSO is known and so the linker will generate a relative
members should only export the function foo and the de- relocation.
structor(s) of baz and baz::s. An export map to achieve
this could look like this: For function calls the result is often as good as it gets.
The code generated by the compiler for a PC-relative
jump and a jump through the PLT is identical. It is just
the code which is called (the target function versus the
{ code in the PLT) which makes the difference. The code
global:
is only not optimal in one case: if the function call is the
extern "C++" {
foo*;
only reason the PIC register is loaded. For a call to a local
baz::baz*; function this is not necessary and loading the PIC is just
baz::s* a waste of time and code.
};
local: *; To summarize, for variables the use of symbol maps cre-
}; ates larger and less efficient code, adds an entry in the
GOT, and adds a relative relocation. For functions the
generated code sometimes contains unnecessary loads of
the PIC. One normal relocation is converted into a rel-
The use of extern "C++" tells the linker to match the ative relocation and one PLT entry is removed. This is
following patterns with demangled C++ names. The first one relative relocation worse than the previous methods.
entry foo* matches the first global function in the ex- These deficiencies are the reason why it is much prefer-
ample. The second entry matches the constructor(s) of able to tell the compiler what is going on since after the
baz and the third entry matches the function baz::s. compiler finished its work certain decisions cannot be re-
Note that patterns are used in all cases. This is necessary verted anymore.
since foo, baz::baz, and baz::s are not the complete
names. The function parameter are also encoded in the
mangled name and must be matched. It is not possible to 2.2.6 Libtool’s -export-symbols
match complete demangled C++ names since the current
linker implementation refuses to allow non-alphanumeric The fourth method to restrict symbol export is the least
characters. Using pattern might have unwanted effects. desirable of them. It is the one used by the GNU Libtool
If there is another member function in baz starting with program when the -export-symbols option is used.
the letter ‘s’ it will also be exported. And one last odd- This option is used to pass to Libtool the name of a file
ity should be mentioned: currently the linker requires that which contains the names of all the symbols which should
there is no semicolon after the last entry in the C++ block. be exported, one per line. The Libtool command line
might look like this:
Using export maps seems like a very desirable solution.
The sources do not have to be made less readable us-
ing attribute declarations or eventually pragmas. All the
knowledge of the ABI is kept locally in the export map $ libtool --mode=link gcc -o libfoo.la \
file. But this process has one fundamental problem: ex- foo.lo -export-symbols=foo.sym
actly because the sources are not modified the final code
is not optimal. The linker is used only after the compiler
already did its work and the once generated code cannot
be optimized significantly. The file foo.sym would contain the list of exported sym-
bols. foo.lo is the special kind of object files Libtool
In our running example the compiler must generate the generates. For more information on this and other strange
code for the next function under the worst case sce- details from the command line consult the Libtool man-
nario assumption that the variable last is exported. This ual.
means the code sequence using @GOTOFF which was men-
tioned before cannot be generated. Instead the normal Interesting for us here is the code the linker produces us-
two instruction sequence using @GOT must be generated. ing this method. For the GNU linker Libtool converts the
-export-symbols option into the completely useless
This is what the linker will see when it gets instructed to -retain-symbols-file option. This option instructs
hide the symbol last. The linker will not touch the ac- the linker to prune the normal symbol tables, not the dy-
tual code. Code relaxation here would require substantial namic symbol table. The normal symbol table will con-
analysis of the following code which is in theory possi- tain only the symbols named in the export list file plus the
ble but not implemented. But the linker will not generate special STT SECTION symbols which might be needed in
the normal R 386 GLOB DAT relocation either. Since the relocations. All local symbols are gone. The problem is
symbol is not exported no interposition is allowed. The that the dynamic symbol table is not touched at all and
position of the local definition relative to the start of the this is the table which is actually used at runtime.

Ulrich Drepper Version 4.1.2 23

The effect of the using libtool this way is that pro- as well as next we could use code like this:
grams reading the normal symbol table (for instance nm)
do not find any symbols but those listed in the export
list. And that is it. There are no runtime effects. Neither
static int last;
have any symbols been made unavailable for the dynamic
linker, nor have any normal relocations been converted
static int next_int (void) {
into relative relocations. return ++last;
}
The only reason this method is mentioned here is that
there is hope libtool will learn about converting the int next (void) {
export lists into the anonymous version maps we have return next_int ();
seen in the previous section when the GNU linker is used. }
At that point libtool will become useful. Until then
relying on its -export-symbols option is misleading int index (int scale) {
at best. return next_int () << scale;
}

2.2.7 Avoid Using Exported Symbols

The function next is now a simple wrapper around next int.

In some situations it might not be desirable to avoid ex- All calls to next int are recognized by the compiler as
porting a symbol but at the same time all local references calls to a local function since next int, unlike next, is
should use the local definition. This also means that the defined with static. Therefore no PLT entries is used
uses of the symbols is cheaper since the less general code for the call in index.
sequences can be used. This is a subset of the problem
discussed so far. A solution needs a different approach The drawback of this method is that additional code is re-
since so far we achieved the better code by not exporting quired (the code for the new next function) and that call-
only. ing next also minimally slower than necessary at run-
time. As a fallback solution, in case no other method
Since a symbol cannot be exported and not-exported at works, this is better than nothing.
the same time the basic approach is to use two names for
the same variable or function. The two names then can
be treated differently. There are multiple possibilities to Using Aliases Introducing two names without adding
create two names, varying in efficiency and effort. code can be achieved by creating aliases for existing ob-
jects. Support for this is included in gcc; this does not
At this point it is necessary to add a warning. By per- only include the creation of the alias, gcc also knows the
forming this optimization the semantics of the program type for the alias and can perform appropriate tests when
changes since the optimization interferes with the sym- the alias is used. The here goal is therefore to create an
bol lookup rules. It is now possible to use more than one alias and tell gcc and/or the linker to not export the sym-
symbol with a given name in the program. Code out- bol. I.e., we apply the same techniques described in the
side the DSO might find a definition of a symbol some- previous sections now to an alias. The only difference is
where else while the code in the DSO always uses the that defining an alias as static will not work. The best
local definition. This might lead to funny results. Of- method therefore is to use visibility attributes. The other
ten it is acceptable since multiple definitions are not al- previously discussed methods will also work but we do
lowed. A related issue is that one rule of ISO C can be not go into the details for them here.
violated by this. ISO C says that functions are identified
by their names (identifiers) and that comparing the func- If in our example we want to export both last and next
tion pointers one can test for equality. The ELF imple- we can rewrite the example like this:
mentation works hard to make sure this rule is normally
obeyed. When forcing the use of local symbols code in-
side and outside the DSO might find different definitions
for a given name and therefore the pointers do not match. int last;
It is important to always consider these side effects before extern __typeof (last) last_int
performing the optimization. __attribute ((alias ("last"),
visibility ("hidden")));

int next (void) {

Wrapper Functions Only applicable to functions, us- return ++last_int;
ing wrappers (i.e. alternative entry points) is the most }
portable but also most costly way to solve the problem. extern __typeof (next) next_int
If in our example code we would want to export index __attribute ((alias ("next"),

24 Version 4.1.2 How To Write Shared Libraries

 visibility ("hidden"))); rect symbol name and therefore the alias definition will
fail. For this reason it is mandatory to create alias only of
int index (int scale) { non-static functions and variables.
return next_int () << scale;
} For C++ programs defining aliases we also are also chal-
lenged by names. The problem is that the alias attribute
requires the assembler name of the defined symbol as a
string parameter. For C++ code this means the mangled
This is quite a collection of non-standard gcc extensions name. For simple C++ function we manage to get along
to the C language so it might need some explanation. The with the same trick used in the C example.
actual definitions of all three objects are the same as in the
original code. All these objects are exported. The differ-
ence in the definitions is that next is using the internal
alias last int instead of last and similarly for index int
and next. What looks like two declarations is the mech- add (int a, int b)
anism by which gcc is told about the aliases. It is basi- {
cally an extern declaration of an object with the same return a + b;
}
type (we use here typeof to ensure that) which has an
extern __typeof (add) add_int
alias added. The alias attribute names the object this
__attribute ((alias ("_Z3addii"),
is an alias of. visibility ("hidden")));

To achieve the results we want, namely that the aliases are

not exported and that gcc gets told about this, we have
to add the hidden visibility attribute. Looking back at There are only two tricky parts. The first is finding the
sections 2.2.2 and 2.2.3 it should be easy to see that the correct mangled name. For the locally used compiler it is
use of this attribute is equivalent. quite easy to determine the name, just compile the code
without the alias definition and look at the symbol table
If the visibility attributes cannot be used for some reason of the generated file. Name mangling is unfortunately
almost the same code should be used, only leaving out traditionally not well standardized. There exist several
different name mangling schemes which means the alias
, visibility ("hidden") string would have to be adjusted to the compiler which is
used for the compilation.
This will create a normal alias with the same scope as the
original symbol. Using export maps the alias can then The second problem is the use of typeof if the func-
be hidden. The resulting binary will not use the efficient tion name is overloaded. In this case the compiler does
code sequences (see section 2.2.5) but the local definition not know which of the potentially many versions of the
will always be used. function is meant and it bails out.

An attentive reader might suggest that it is possible to

avoid some of the complications by writing the code for DF SYMBOLIC The original designers of the ELF for-
next like this: mat considered the possibility that preferring local def-
initions might be useful. They have included a mecha-
nism which can enforce this. If the DF SYMBOLIC flag
is set in the DT FLAGS entry of the dynamic section (or
static int next_int (void) {
in older ELF binaries: if the dynamic section contains
return ++last_int;
}
an DT SYMBOLIC entry) the dynamic linker has to prefer
extern __typeof (next_int) next local definitions.
__attribute ((alias ("next_int")));
This approach has numerous disadvantages. First, all in-
terfaces are affected. The other approaches discussed
here have a per-interface granularity. Treating all inter-
As a static definition, next int is not exported and el- faces like this is normally not the right way. The second
igible for inlining, while next is defined as extern and disadvantage is that the compiler does not get told about
therefore exported. Even though this sometimes works the use of local symbols and therefore cannot optimize
there is no guarantee it does all the time. The compiler the uses, just as if export maps would be used. And what
is allowed to use arbitrary symbol names for static is even worse, calls to local functions still use the PLT en-
functions and variables since the names are not part of tries. The PLT and GOT entries are still created and the
the ABI of the object file. This is necessary in some jump is indirect. This might be useful in some situations
situations to avoid name clashes. The result is that the (e.g., when using LD PROFILE) but usually means a big,
alias("next int") part might fail to provide the cor- missed opportunity for optimization.

Ulrich Drepper Version 4.1.2 25

Finally the third problem is that the lookup scope is changed any role at run-time.
in a way which can lead to using unexpected dependen-
cies. DF SYMBOLIC effectively puts the own object in The wrapper class will have to redefine all non-virtual
the first spot of its own lookup scope so that there are member functions of the class it is helping to export. This
a number of other DSO which are seen before the depen- requires some work and it might add run-time costs by
dencies. This is nothing new but the fact that the DSO an additional function call, the one to the wrapper func-
marked with DF SYMBOLIC is in an unusual place can tion. Note that defining those functions inline will not
cause unexpected versions from being picked up. help since then the reference to the original, long name
is reintroduced. The only way to avoid the extra call is
The advice here is to never use DF SYMBOLIC. It does to define appropriate aliases, which might prove cumber-
not improve the code, forces all symbols to be treated the some.
same, and can cause problems in symbol lookup. It is
mentioned here only for completeness and as a warning. Shortening symbol names can be considered a micro-
optimization and certainly should not be performed pre-
2.3 Shortening Symbol Names maturely. When keeping this optimization in mind dur-
ing the development it might be easy to implement and
the possible benefits can be big. Memory operations are
The description of the ELF symbol lookup algorithm shows slow and if the number of bytes which have to be loaded
that one of the cost factors for the lookup is length of can be reduced this definitely has measurable results.
the symbols involved. For successful lookups the entire
string has to be matched and unsuccessful lookups re- 2.4 Choosing the Right Type
quire matching the common prefix of the involved strings.

The flat namespace of the C programming environment

The selection of the right type can have significant im-
makes following the guideline to use short names easy.
pact on the performs, startup time, and size of a program.
The names the programmer uses are directly mapped to
Most of the time it seems obvious what the right type
names in the ELF file. The same is true for some other
is but alternatives are sometimes available and in other
programming environments such as traditional Pascal and
cases it might be preferable to rearrange code slightly.
FORTRAN.
This section will provide a few concrete examples which
by no means are meant to be a complete representation of
Programming environments with more sophisticated sym-
all the cases which can be optimized.
bol handling use name mangling. The most prominent
programming language in this class is C++. The sym-
bol string of a function consists beside the function name 2.4.1 Pointers vs. Arrays
also of a description of the parameter list, the classes the
function is a member of and the namespaces the class
In some situations there is little or no difference between
or function is defined in. This can lead to enormously
pointers and arrays in C. The prototypes
long symbol names. Names of more than 1,000 charac-
ters have been sighted in the wild. Ada names tend to get
very long because of the package namespace.
void scale (int arr[10], int factor)
The object and namespace model in C++ is used to man-
age the complexity of large projects and to facilitate code
reuse. Therefore it is desirable keep symbol names un-
modified during the development process. But once a and
program is to be deployed the long names become a nui-
sance. This is when a person can step in and shorten the
names.
void scale (int *arr, int factor)
In C++ the most critical classes are those with templates
and/or deeply nested namespaces and class definitions. If
such classes are part of the interface of a DSO the pro-
grammer should make a change. A shorter name for a are in fact mostly equivalent. So people got the impres-
class can be introduced by deriving publically a new class sion that there is never a difference and one often finds
from the class with the long name. The definition could code like this:
be in the global scope to avoid the namespace part of the
mangled name. The symbols associated with this new
class can be exported, the original class’ names are not.
This does not remove the names of the original class from char *str = "some string";
the non-dynamic symbol table but this table does not play

26 Version 4.1.2 How To Write Shared Libraries

This is correct and meaningful in some situations. A vari-
able str is created with an initial value being a pointer
to a string. This specific piece of code compiles fine with const char s1[] = "some string";
some compilers but will generate a warning when com- const char s2[] = "string";
piled with gcc. More on that in the next section.

The point to be made here is that the use of a variable

in this situation is often unnecessary. There might not be In this case only the string "some string" has to be
an assignment to str (note: not the string, the pointer stored in the read-only data segment. The symbol s2 can
variable). The value could be used only in I/O, string be a reference to the fifth character of the longer string.
generation, string comparison or whatever.
To make this possible the compiler has to emit the string
If this is the case the code above is not optimal and wastes data in specially marked section. The sections are marked
resources. All that would be needed is the string itself. A with the flags SHF MERGE and SHF STRINGS.
better definition would therefore be:
Not all strings can be handled, though. If a string con-
tains an explicit NUL byte, as opposed to the implicit
one at the end of the string, the string cannot be placed
char str[] = "some string";
in mergeable section. Since the linker’s algorithms use
the NUL byte to find the end of the string the rest of the
input string would be discarded. It is therefore desirable
This is something completely different than the code be- to avoid strings with explicit NUL bytes.
fore. Here str is a name for a sequence of bytes which
contains initially the sequence "some string". By rewrit- 2.4.3 Arrays of Data Pointers
ing code in this way whenever it is possible we save one
pointer variable in the non-sharable data segment, and
one relative relocation to initialize the variable with a Some data structure designs which work perfectly well
pointer to the string. Eventually the compiler is able to in application code add significant costs when used in
generate better code since it is known that the value of DSOs. This is especially true for arrays of pointers. One
str can never change (the bytes pointed to by str can example which shows the dilemma can be met frequently
change). in well-designed library interface. A set of interfaces re-
turns error number which can be converted using another
function into strings. The code might look like this:
2.4.2 Forever const

One nit still exists with the result in the last section: the
string is modifiable. Very often the string will never be static const char *msgs[] = {
modified. In such a case the unsharable data segment is [ERR1] = "message for err1",
unnecessarily big. [ERR2] = "message for err2",
[ERR3] = "message for err3"
};

const char str[] = "some string"; const char *errstr (int nr) {
return msgs[nr];
}

After adding the const keyword the compiler is able

to move the string in sharable read-only memory. This
not only improves the program’s resource use and startup The problematic piece is the definition of msgs. msgs is
speed, it also allows to catch mistakes like writing into as defined here a variable placed in non-sharable, writable
this string. memory. It is initialized to point to three strings in read-
only memory (that part is fine). Even if the definition
But that is not all. Modern gcc and linker versions can would be written as
work together to perform cross-object optimizations. I.e.,
strings which appear in more than one object file appear static const char *const msgs[] = {
only once in the final output. And even more: some link-
ers perform suffix optimizations, something which is pos- (note the addition const) this would not change this (but
sible with the string representation used in C. For this it it opens up some other opportunities, see 2.6). The com-
is necessary to realize that a string, which is the back- piler still would have the place the variable in writable
part of a longer string (including the NUL byte), can be memory. The reason are three relative relocation which
represented by the bytes from the longer string. modify the content of the array after loading. The total

Ulrich Drepper Version 4.1.2 27

cost for this code is three words of data in writable mem- };
ory and three relocations modifying this data in addition
to the memory for the strings themselves. const char *errstr (int nr) {
return msgstr + msgidx[nr];
Whenever a variable, array, structure, or union, contains }
a pointer, the definition of an initialized variable requires
relocations which in turn requires the variable to be placed
in writable memory. This along with the increased startup
time due to processing of the relocations is not acceptable The content of both arrays in this code is constant at
for code which is used only in error cases. compile time. The references to msgstr and msgidx
in errstr also do not need relocations since the defini-
For a simple case as the example above a solution entirely tions are known to be local. The cost of this code include
within C can be used by rewriting the array definition like three size t words in read-only memory in addition to
this: the memory for the strings. I.e., we lost all the reloca-
tions (and therefore startup costs) and moved the array
from writable to read-only memory. In this respect the
code above is optimal.
static const char msgs[][17] = {
[ERR1] = "message for err1", For a more elaborate and less error-prone method of con-
[ERR2] = "message for err2", structing such tables is appendix B. The presented code
[ERR3] = "message for err3"
does not require the programmer to duplicate strings which
};
must be kept in sync.

2.4.4 Arrays of Function Pointers

The result of this code is optimal. The array msgs is
placed entirely in read-only memory since it contains no
pointer. The C code does not have to be rewritten. The The situation for function pointers is very similar to that
drawback of this solution is that it is not always applica- of data pointers. If a pointer to a function is used in the
ble. If the strings have different lengths it would mean initialization of a global variable the variable the result
wasting quite a bit of memory since the second dimen- gets written to must be writable and non-sharable. For
sion of the array has to be that of the length of the longest locally defined functions we get a relative relocation and
string plus one. The waste gets even bigger if the values for functions undefined in the DSO a normal relocation
ERR0, ERR1, and ERR2 are not consecutive and/or do not which is not lazily performed. The question is how to
start with zero. Every missing entry would mean, in this avoid the writable variable and the relocations. Unfortu-
case, 17 unused bytes. nately there is no generally accepted single answer. All
we can do here is to propose a few solutions. Our exam-
There are other methods available for case which cannot ple code for this section is this:
be handled as the example above but none without major
code rewrite.10 One possible solution for the problem is
the following. This code is not as elegant as the original
code but it is still maintainable. Ideally there should be a static int a0 (int a) { return a + 0; }
tool which generates from a description of the strings the static int a1 (int a) { return a + 1; }
appropriate data structures. This can be done with a few static int a2 (int a) { return a + 2; }
lines
static int (*fps[]) (int) = {
[0] = a0,
[1] = a1,
static const char msgstr[] = [2] = a2
"message for err1\0" };
"message for err2\0"
"message for err3"; int add (int a, int b) {
return fps[b] (a);
static const size_t msgidx[] = { }
0,
sizeof ("message for err1"),
sizeof ("message for err1")
+ sizeof ("message for err2") A solution for this problem must inevitably be differ-
10 Ifwe would write assembler code we could store offsets relative to
ent from what we did for strings where we combined all
a point in the DSO and add the absolute address of the reference point strings into one. We can do this for functions as well but
when using the array elements. This is unfortunately not possible in C. it will look different:

28 Version 4.1.2 How To Write Shared Libraries

 cations in a DSO and which require that the array labels
is writable and placed in non-sharable memory.
int add (int a, int b) {
switch (b) { The above code in principal is an implementation of a
case 0: switch statement. The difference is that the compiler
return a + 0; never stores absolute addresses which would need relo-
case 1: cations of position-independent code is generated. In-
return a + 1;
stead the addresses are computed relative to the PIC ad-
case 2:
dress, producing a constant offset. This offset then has
return a + 2;
} to be added to the PIC register value which is a minimal
} amount of extra work. To optimize the code above a sim-
ilar scheme but be used.

Inlining the code as in the code above should certainly be

the preferred solution. The compiler never generates re- int add (int a, int b) {
locations for the implementation of a switch statement static const int offsets[] = {
and therefore the whole code does not need any reloca- &&a0-&&a0, &&a1-&&a0, &&a2-&&a0
};
tion.
goto *(&&a0 + offsets[b]);
a0:
Inlining makes sense even if the inlined function are much return a + 0;
larger. Compilers do not have problems with large func- a1:
tions and might even be able to optimize better. The return a + 1;
problem is only the programmer. The original code was a2:
clean and intentionally written using function pointers. return a + 2;
The transformed code might be much less readable. This }
makes this kind of optimization one which is not per-
formed until the code is tested and does not need much
maintenance anymore.
Since we do not have direct access to the PIC register at
A similar problem, though it (unfortunately) is rather rare compile-time and cannot express the computations of the
today, arises for the use of computed gotos, a gcc ex- offsets we have to find another base address. In the code
tension for C. Computed gotos can be very useful in above it is simply one of the target addresses, a0. The
computer-generated code and in highly optimized code.11 array offsets is in this case really constant and placed
The previous example using computed gotos might look in read-only memory since all the offsets are known once
like this: the compiler finished generating code for the function.
We now have relative addresses, no relocations are nec-
essary. The type used for offsets might have to be ad-
justed. If the differences are too large (only really pos-
int add (int a, int b) { sible for 64-bit architectures, and then only for tremen-
static const void *labels[] = { dously large functions) the type might have to be changed
&&a0, &&a1, &&a2 to ssize t or something equivalent. In the other direc-
};
tion, if it is known that the offsets would fit in a variable
goto *labels[b];
of type short or signed char, these types might be
a0:
return a + 0; used to save some memory.
a1:
return a + 1; 2.4.5 C++ Virtual Function Tables
a2:
return a + 2;
} Virtual function tables, generated for C++ classes with
member functions tagged with virtual, are a special
case. The tables normally involve function pointer which
cause, as seen above, the linker to create relocations. It
How the code works should be obvious. The array labels is also worthwhile looking at the runtime costs of virtual
contains pointers to the places in the function where the functions compared to intra- and inter-DSO calls. But
labels are placed and the gotos jumps to the place picked first the relocation issues.
out of the array. The problem with this code is that the ar-
ray contains absolute address which require relative relo- Usually the virtual function table consists of an array of
11 Interested readers might want to look at the vfprintf implemen- function pointers or function descriptors. These represen-
tation in the GNU libc. tations have in common that for every slot in the virtual

Ulrich Drepper Version 4.1.2 29

function table there must be one relocation and all the virtual int virfunc () const;
relocations have to be resolved at startup-time. There- int virfunc_do () const;
fore having larger a number of virtual function tables or };
virtual function tables with many entries impact startup
time. /* In the source code file. */
virtual int foo::virfunc () const
One other implication is noteworthy. Even though at run- { return virfunc_do (); }
int foo::virfunc_do () const
time normally only one virtual function table is used (since
{ ...do something... }
the name is the same the first is in the lookup scope is
used) all tables must be initialized. The dynamic linker
has no possibility to determine whether the table will be
used at some point or not and therefore cannot avoid ini-
In this case the real implementation of the function is in
tializing it. Having exactly one virtual function table def-
virfunc do and the virtual function just calls it. There
inition in all the participating DSOs is therefore not only
is no need for the user to call the virtual function di-
useful for space conservation reasons.
rectly as in the function bar above since virfunc do
can be called instead. Therefore the linker version script
The cost of each relocation depends on whether the func-
could hide the symbol for the virtual function and export
tion being referenced is defined locally and whether it
virfunc do to be called from other DSOs. If the user
is exported or not. Only for functions which are explic-
still calls the virtual function the linker will not be able to
itly hidden by using the visibility attributes or a version
find a definition and the programmer has to know that this
script can the linker use relative relocations which can
means she has to rewrite the code to use virfunc do.
be processed quickly. Otherwise relocations with symbol
This makes using the class and the DSO a bit more com-
lookups have to be used. Using the visibility attributes
plex.
was mentioned as a possibility to get relative relocations.
But since the virtual function table and the instantiated
The consequence of hiding the virtual function is that the
member functions are generated by the compiler adding
virtual function table slot for virfunc can be handled
an attribute cannot be done without some dirty tricks. So
with a relative relocation. This is a big gain not only be-
using linker version scripts is really the only possibility.
cause this relocation type is much faster to handle. Since
virtual function tables contain data references the relo-
But even this is often not possible. The virtual functions
cation of the virtual function table slots must happen at
can be called not only through the virtual function table
startup time.
but also directly if the compiler can determine the exact
type of an C++ object. Therefore virtual function in most
The improved example above assumes that direct calls
cases have to be exported from the DSO. For instance,
are more frequent than calls through the virtual func-
the virtual function in the following example is called di-
tion table since there is additional overhead (one extra
rectly.
function call) involved when calling virfunc. If this as-
sumption is wrong and calls through the virtual function
table are more frequent, then the implementations of the
struct foo { two functions can be swapped.
virtual int virfunc () const;
}; In summary, the number and size of virtual function ta-
foo var; bles should be kept down since it directly impacts startup
int bar () { return var.virfunc (); } time behavior. If virtual functions cannot be avoided the
implementations of the functions should not be exported.

The reason is that var is known to have type foo and 2.5 Improving Generated Code
not a derived type from which the virtual function table
is used. If the class foo is instantiated in another DSO
not only the virtual function table has to be exported by On most platforms the code generated for DSOs differs
that DSO, but also the virtual function virfunc. from code generated for applications. The code in DSOs
needs to be relocatable while application code can usu-
If a tiny runtime overhead is acceptable the virtual func- ally assume a fixed load address. This inevitably means
tion and the externally usable function interface should that the code in DSOs is slightly slower and probably
be separated. Something like this: larger than application code. Sometimes this additional
overhead can be measured. Small, often called functions
fall into this category. This section shows some problem
cases of code in DSOs and ways to avoid them.
/* In the header. */
struct foo { In the preceding text we have seen that for IA-32 a func-

30 Version 4.1.2 How To Write Shared Libraries

tion accessing a global variable has to load determine the loaded before making the call. The result is that for our
address of the GOT to use the @GOTOFF operation. As- running example the generated code might look like this:
suming this C code

getfoo:
static int foo; addl r14=@gprel(foo),gp;;
int getfoo (void) ld4 r8=[r14]
{ return foo; } br.ret.sptk.many b0

the compiler might end up creating code like this: If the caller knows that the called function uses the same
gp value, it can avoid the loading of gp. IA-32 is really a
special case, but still a very common one. So it is appro-
priate to look for a solution.
getfoo:
call 1f Any solution must avoid the PIC register entirely. We
1: popl %ecx propose two possible ways to improve the situation. First,
addl _GLOBAL_OFFSET_TABLE_[.-1b],%ecx
do not use position-independent code. This will generate
movl foo@GOTOFF(%ecx),%eax
ret
code like

getfoo:
The actual variable access is overshadowed by the over-
movl foo,%eax
head to do so. Loading the GOT address into the %ecx ret
register takes three instructions. What if this function
is called very often? Even worse: what if the function
getfoo would be defined static or hidden and no pointer
to it are ever available? In this case the caller might al- The drawback is that the resulting binary will have text
ready have computed the GOT address; at least on IA-32 relocations. The page on which this code resides will
the GOT address is the same for all functions in the DSO not be sharable, the memory subsystem is more stressed
or executable. The computation of the GOT address in because of this, a runtime relocation is needed, program
foobar would be unnecessary. The key word in this sce- startup is slower because of both these points, and secu-
nario description is “might”. The IA-32 ABI does not rity of the program is impacted. Overall, there is a mea-
require that the caller loads the PIC register. Only if a surable cost associated with not using PIC. This possibil-
function calls uses the PLT do we know that %ebx con- ity should be avoided whenever possible. If the DSO is
tains the GOT address and in this case the call could come question is only used once at the same time (i.e., there are
from any other loaded DSO or the executable. I.e., we re- no additional copies of the same program or other pro-
ally always have to load the GOT address. grams using the DSO) the overhead of the copied page is
not that bad. Only in case the page has to be evacuated
On platforms with better-designed instruction sets the gen- from memory we would see measurable deficits since the
erated code is not bad at all. For example, the x86-64 page cannot simply be discarded, it must be stored in the
version could look like this: disk swap storage.

The second proposed solution has a bigger impact on the

whole code. Assume this extended example:
getfoo:
movl foo(%rip),%eax
ret
static int foo;
static int bar;
int getfoo (void)
The x86-64 architecture provides a PC-relative data ad- { return foo; }
dressing mode which is extremely helpful in situations int getboth (void)
like this. { return bar+getfoo(); }

Another possible optimization is to require the caller to

load the PIC register. On IA-64 the gp register is used
for this purpose. Each function pointer consist of a pair If this code gets translated as is, both functions will load
function address and gp value. The gp value has to be the GOT address to access the global variables. This can

Ulrich Drepper Version 4.1.2 31

be avoided by putting all variables in a struct and passing setuid to drop a process’ privileges which is redirected
the address of the struct to the functions which can use it. to perhaps getpid. The attacker could therefore keep
For instance, the above code could be rewritten as: the raised priviledges and later cause greater harm.

This kind of attack would not be possible if the data GOT

and PLT could not be modified by the user program. For
static struct globals { some platforms, like IA-32, the PLT is already read-only.
int foo;
But the GOT must be modifiable at runtime. The dy-
int bar;
namic linker is an ordinary part of the program and it
} globals;
static int intfoo (struct globals *g) is therefore not possible to require the GOT in a mem-
{ return g->foo; } ory region which is writable by the dynamic linker but
int getfoo (void) not the rest of the application. Another possibility would
{ return intfoo(&globals); } be to have the dynamic linker change the access permis-
int getboth (void) sions for the memory pages containing the GOT and PLT
{ return globals.bar+intfoo(&globals); } whenever it has to change a value. The required calls to
mprotect are prohibitively expensive, ruling out this so-
lution for any system which aims to be performing well.

The code generated for this example does not compute At this point we should remember how the dynamic linker
the GOT address twice for each call to getboth. The works and how the GOT is used. Each GOT entry be-
function intfoo uses the provided pointer and does not longs to a certain symbol and depending on how the sym-
need the GOT address. To preserve the semantics of the bol is used, the dynamic linker will perform the relo-
first code this additional function had to be introduced; cation at startup time or on demand when the symbol
it is now merely a wrapper around intfoo. If it is pos- is used. Of interest here are the relocations of the first
sible to write the sources for a DSO to have all global group. We know exactly when all non-lazy relocation
variables in a structure and pass the additional parameter are performed. So we could change the access permis-
to all internal functions, then the benefit on IA-32 can be sion of the part of the GOT which is modified at startup
big. time to forbid write access after the relocations are done.
Creating objects this way is enabled by the -z relro
But it must be kept in mind that the code generated for the linker option. The linker is instructed to move the sec-
changed example is worse than what would be created for tions, which are only modified by relocations, onto sep-
the original on most other architectures. As can be seen, arate memory page and emit a new program header en-
in the x86-64 case the extra parameter to intfoo would try PT GNU RELRO to point the dynamic linker to these
be pure overhead since we can access global variables special pages. At runtime the dynamic linker can then
without the GOT value. On IA-64 marking getfoo as remove the write access to these pages after it is done.
hidden would allow to avoid the PLT and therefore the gp
register is not reloaded during the call to getfoo. Again, This is only a partial solution but already very useful.
the parameter is pure overhead. For this reason it is ques- By using the -z now linker option it is possible to dis-
tionable whether this IA-32 specific optimization should able all lazy relocation at the expense of increased startup
ever be performed. If IA-32 is the by far most important costs and make all relocations eligible for this special
platform it might be an option. treatment. For long-running applications which are secu-
rity relevant this is definitely attractive: the startup costs
2.6 Increasing Security should not weigh in as much as the gained security. Also,
if the application and DSOs are written well, they avoid
relocations. For instance, the DSOs in the GNU C library
The explanation of ELF so far have shown the critical im- are all build with -z relro and -z now.
portance of the GOT and PLT data structures used at run-
time by the dynamic linker. Since these data structures The GOT and PLT are not the only parts of the applica-
are used to direct memory access and function calls they tion which benefit from this feature. In section 2.4.2 we
are also a security liability. If a program error provides have seen that adding as many const to a data object def-
an attacker with the possibility to overwrite a single word inition as possible has benefits when accessing the data.
in the address space with a value of his choosing, then But there is more. Consider the following code:
targetting a specific GOT entry is a worthwhile goal. For
some architectures, a changed GOT value might redirect
a call to a function, which is called through the PLT, to
an arbitrary other place. Similarly, if the PLT is modified const char *msgs1[] = {
in relocations and therefore writable, that memory could "one", "two", "three"
be modified. };
const char *const msgs2[] = {
An example with security relevance could be a call to "one", "two", "three"

32 Version 4.1.2 How To Write Shared Libraries

}; any code. By setting the value of the environment vari-
able LD PROFILE to the shared object name (SONAME)
of a DSO, all uses of interfaces defined in this DSO which
go through a PLT are recorded. In addition time-based
It has been explained that neither array can be moved sampling happens as well and all samples which happen
into the read-only, and therefore shared, data section if in the profiled DSO are recorded. This is all similar to the
the file is linked into a DSO. The addresses of the strings gprof-based profiling which is available on Unix sys-
pointed to by the elements are known only at runtime. tems for a long time.
Once the addresses are filled in, though, the elements of
msgs2 must never be modified again since the array it- The LD PROFILE variable can be set for arbitrarily many
self is defined as const. Therefore gcc stores the array applications running at the same time. All profiling activ-
msgs1 in the .data section while msgs2 is stored into ity is recorded in real-time in the same output file which is
a section called .data.rel. This .data.rel section is usually stored in /var/tmp. Without having to stop any
just like .data but the dynamic linker could take away or all applications currently profiled, the current status
write access after the relocations are done. The previ- can be examined by running the sprof program. The op-
ously described handling of the GOT is just a special case tions this program understands are similar to those gprof
of exactly this feature. Adding as many const as possi- understand. There are basically three report types, call
ble together with the -z relro linker option therefore pairs, flat profiles, and graph profiles which can all be
protects even program data. This might even catch bugs requested at once.
where code incorrectly modifies data declared as const.
How to use these profile results is specific to the applica-
A completely different issue, but still worth mentioning tion. It is easy to locate often called functions and, given
in the context of security, are text relocations. Generat- sufficient runtime, functions inside the DSO which run
ing DSOs so that text relocations are necessary (see sec- for a long time. Everybody who used gprof should feel
tion 2) means that the dynamic linker has to make mem- right at home although DSO profiling has its limitations.
ory pages, which are otherwise read-only, temporarily
writable. The period in which the pages are writable is The call data is accumulated by intercepting the calls
usually brief, only until all non-lazy relocations for the made through the PLT. The dynamic linker can do this
object are handled. But even this brief period could be without the application noticing it. But if the PLT is not
exploited by an attacker. In a malicious attack code re- used, data is not recorded. This is the case if an appli-
gions could be overwritten with code of the attacker’s cation looks a symbol up using dlsym and then calls it
choice and the program will execute the code blindly if it directly. This might make the profiling data useless.
reaches those addresses.
The solution for this problem is to use the DL CALL FCT
During the program startup period this is not possible macro the GNU C library’s <dlfcn.h> header defines.
since there is no other thread available which could per- If a looked-up function pointer fctp has been used like
form the attack while the pages are writable. The same is this
not true if later, when the program already executes nor-
mal code and might have start threads, some DSOs are foo = fctp (arg1, arg2);
loaded dynamically with dlopen. For this reason cre-
ating DSOs with text relocation means unnecessarily in- or if you prefer
creasing the security problems of the system.
foo = (*fctp) (arg1, arg2);
Much more critical is that with Security Enhanced Linux
(SELinux) text relocations become a big liability. By de- the code should be rewritten to look like this:
fault, the SELinux extensions prevent making executable
data pages writable. Since the kernel cannot distinguish foo = DL CALL FCT (fctp, (arg1, arg2));
between the dynamic linker doing this and the program
or an attacker making the change, the abilities of the dy- The DL CALL FCT macro contains the necessary magic
namic linker are also restricted. The only ways out are to record the calling of the function pointed to by fctp.
to grant the program permission to change the protection If the DSO which contains the symbol is not profiled
(and therefore give attackers the ability to do the same) or nothing happens. It is therefore safe to always use this
to remove text relocations from all DSOs and PIEs. The macro to call symbols in dynamically loaded DSOs. The
former option should only be used as a last resort. DL CALL FCT macro has a low overhead and so could be
used unconditionally but for production code it is proba-
2.7 Simple Profiling of Interface Usage bly best to avoid it.

The dynamic linker in the GNU C library allows easy

profiling of the way a DSO is used without recompiling

Ulrich Drepper Version 4.1.2 33

3 Maintaining APIs and ABIs ments therefore is that the parameters of a function do not
change. This brings up an interesting point: in C++ this
is ensured automatically. Functions incorporate in their
When writing DSOs which are used as resources in mul-
mangled names the parameter types. This means that any
tiple projects mastering the technical aspects of writing
change in the signature of a function will result in link-
optimized DSOs is only part of what is needed. Main-
time and run-time errors and therefore can be detected
taining the programming interface (API) and the binary
easily. This is not the case for variables; their mangled
interface (ABI) play an even more important role in a suc-
names only contain the namespace part. Another good
cessful project. Without API and ABI stability the DSO
reason to not export variables as part of the API.
would be a burden to use or even unusable. In this sec-
tion we will introduce a number of rules which increase
3.2 Defining Stability
the chance of success for projects which provides inter-
faces for other projects. We are talking specifically about
library implementations in DSOs but most rules can be Having said that stability of the ABI is the highest goal
transferred to projects of any kind. of DSO maintenance, it is now time to define what sta-
bility means. This might be surprising for some readers
3.1 What are APIs and ABIs? as a naı̈ve view of the problem might be that everything
which worked before has to continue working in the fu-
ture. Everybody who tried this before will see a problem
DSOs are used both at compile time and at run-time. At with it.
compile time the linker tries to satisfy undefined refer-
ences from the definitions in the DSO. The linker then Requiring everything to continue to be possible in future
associates the reference with the definition in the DSO. In releases would mean that even programs, which use in-
ELF objects this reference is not explicitly present since terfaces in an undocumented way, have to be supported.
the symbol lookup allows finding different definitions at Almost all non-trivial function interfaces allow parame-
run-time. But the reference is marked to be satisfied. At ters to specify values which are outside the documented
run-time the program can rely on the fact that a defini- interface and most interfaces are influenced by side ef-
tion is present. If this would not be the case something fects from other functions or previous calls. Requiring
changed in the DSO between the time of linking the ap- that such uses of an interface are possible in future revi-
plication and the time the application is executed. A sions means that not only the interface but also the im-
change in which a symbol vanishes is usually fatal. In plementation is fixed.
some cases definitions in other DSOs can take over but
this is nothing which one can usually be depended on. As an example assume the implementation of the strtok
A symbol once exported must be available at run-time in function in the C run-time library. The standard requires
the future. that the first call if strtok gets passed a non-NULL first
parameter. But what happens if the first call has a NULL as
The ABI of the DSO comprises the collection of all the the first parameter? In this case the behavior is undefined
definitions which were available for use during the life- (not even implemented-defined in this case). Some im-
time of the DSO. Maintaining ABI compatibility means plementations will in this case simply return NULL since
that no definition, also called interface, gets lost. This is this a common side effect of a possible implementation.
only the easy part, though. But this is not guaranteed. The function call might as
well cause the application to crash. Both are valid im-
For variable definitions it also means that the size and plementations but changing from one to the other in the
structure of the variable does not change in a way the lifetime of a DSO would mean an incompatibility.
application cannot handle. What this actually means de-
pends on the situation. If any code outside the DSO di- The question is: does this really constitute an incompat-
rectly accesses the variable, the accessed part of the struc- ibility? No valid program would ever be affected. Only
ture of the variable must not change. On platforms which programs which are not following the documented inter-
require copy relocations to handle accesses to variables face are affected. And if the change in the implementa-
defined in DSOs in the main application (such as IA-32) tion would mean an improvement in efficiency (accord-
the size of the variable must not change at all. Otherwise ing to whatever measure) this would mean broken appli-
variables might increase in size. cations prevent progress in the implementation of a DSO.
This is not acceptable.
The requirements on function definitions are even harder
to check. The documented semantic of a function must The definition of stability should therefore use the doc-
not change. Defining “semantic” for any non-trivial func- umented interface as the basis. Legitimate uses of in-
tion is not easy, though. In the next section we try to terfaces should not be affected by changes in the imple-
define the requirements of a stable interface. Basically mentation; using interfaces in an undefined way void the
stability means that correct programs which ran in the warranty. The same is true for using completely undoc-
past continue to run in the future. One of the require- umented, internal symbols. Those must not be used at

34 Version 4.1.2 How To Write Shared Libraries

all. While this definition of stability is widely accepted it time and not all possible versioning methods are available
does not mean that avoiding or working around changes on all systems.
introduced by changes to the implementation is wrong. It
just is not necessary if the achievement of stability comes The first method is the oldest and coarsest grained one. It
at a cost which almost always is the case. is implemented by using a different filename for each in-
compatible DSO. In ELF binaries dependencies on DSOs
Rejecting stability of undefined functionality is one thing, are recorded using DT NEEDED entries in the dynamic
but what happens if some documented behavior changes? segment. The string associated with the entry has to be
This is can happen for various reasons: the name of an existing DSO. It is usually taken from the
string associated with the DT SONAME entry in the DSO.
Different names can point to different files which makes
• The implementation contains a bug. Other imple- coexistence of different DSOs possible and easy. But
mentations produce different results and this is what while this method is easy to use and such DSOs can eas-
people interested in cross-platform compatibility ily be produced, it has a major drawback. For every re-
are interested in. The old, broken behavior might leased version of the DSO which contains an incompati-
be useful, too. ble change the SONAME would have to be changed. This
can lead to large numbers of versions of the DSO which
• Similarly, alignment with standards, revisions of each differ only slightly from each other. This wastes
them, or existing practice in other implementations resources especially at run-time when the running appli-
can promise gains in the future and therefore mak- cation need more than one version of the DSO. Another
ing a change is useful. problem is the case when one single application loads
more than one version of the DSO. This is easily pos-
• Functionality of an interface gets extended or re- sible if dependencies (other DSOs the application needs)
duced according to availability of underlying tech- pull in independently these different versions. Since the
nology. For instance, the introduction of more ad- two versions of the versioned DSO do not know of each
vanced harddisk drives can handle requests which other, the results can be catastrophic. The only safe way
previous versions cannot handle and these addi- to handle versioning with filenames is to avoid the de-
tional requests can be exposed through function in- scribed situation completely. This is most probably only
terfaces. possible by updating all binaries right away which means
that effectively no versioning happens. The main advan-
tage of filename versioning is that it works everywhere
Not making the changes can have negative results. Blindly
and can be used as a fallback solution in case no other
changing the code will definitely have negative results.
versioning method is available.
Making the change and still maintaining ABI stability re-
quires the use of versioning.
A second method with finer grained control was devel-
oped by Sun for its Solaris OS. In this method DSOs have
These incompatible changes to a DSO are not the only
internal versioning methods (filename versioning is ob-
changes which can cause problems. Adding a new inter-
viously available on top of them). The internal version-
face does not cause problems for existing applications.
ing allows to make compatible changes to a DSO while
But if a new application uses the new interface it will run
avoiding runtime problems with programs running in en-
into problems if at runtime only the old version of the
vironments which have only the old version of the DSO
DSO, the one without the newly added symbol, is avail-
available. Compatible changes mean adding new inter-
able. The user can detect this by forcing the dynamic
faces or defining additional behavior for existing inter-
linker to perform all relocations at load-time by defin-
faces. Each symbol is associated with a version. The
ing LD BIND NOW to an nonempty value in the environ-
versions in a file are described by a non-cyclical graph
ment before starting the application. The dynamic linker
which forms a hierarchy. If a symbol is associated with a
will abort with an error if an old DSO is used. But forc-
version which has a predecessor it means that the proper-
ing the relocations introduces major performance penal-
ties of the symbol associated with the predecessor version
ties (which is the reason why lazy relocations were in-
are also fulfilled in the successor version. In short: a new
troduced in the first place). Instead the dynamic linker
version is defined for a new release of a DSO whenever
should detect the old DSO version without performing
new features are added. The interfaces which changed
the relocations.
in a compatible way get the new version associated. All
the other interfaces must not change and they keep the
3.3 ABI Versioning version they had in the previous release.

When the linker uses the versioned DSO to satisfy a de-

The term “ABI versioning” refers to the process of asso- pendency it also records the version of the used symbol.
ciating an ABI with a specific version so that if necessary This way it gets for each DSO a list of required versions.
more than one version of the ABI can be present at any This list is recorded in the binary which is produced.
one time. This is no new concept but it was refined over

Ulrich Drepper Version 4.1.2 35

With this information available it is now easy for the dy- If possible, projects implementing generally usable DSOs
namic linker to determine at startup-time whether all the should use symbol versioning from day one. Since the
interfaces the application needs are available in at least same techniques are used for symbol export control this is
the version which was used at link-time. To do this the attractive. Unfortunately this versioning scheme requires
dynamic linker has to go through the list of all required changes in the dynamic linker which are currently only
versions and search in the list of defined versions in the available on Linux and GNU Hurd.12 If this is not possi-
referenced DSOs for a matching entry. If no matching ble use Sun’s internal versioning for compatible changes
entry is found the DSO used at runtime is incompatible (really only applies to Solaris). Otherwise there is no op-
and the program does not start up. tion but to change the SONAME for every release with
incompatible and possible even releases with compati-
It should be noted that the number of versions is much ble changes. But the fact that such limited systems ex-
lower than the number of symbols and generally inde- ist should never make this the only implemented way: if
pendent of it. New versions are introduced only for up- better mechanisms are available they should be used.
dates of the DSO package or an OS release. Therefore
the version matching is a quick process. 3.4 Restricting Exports
While Sun’s extensions help to cope with parts of the
stability problem, the much larger problem remains to One of the reasons changes between revisions of a DSO
be solved: how to handle incompatible changes. Every appear incompatible is that users use internal interfaces
non-trivial DSO will sooner or later in its lifetime require of the DSO. This should never happen and usually the
some incompatible changes. Even changes made to cor- official interfaces are documented and the internal inter-
rect problems fall sometimes into this category. Some faces have special names. Still, many users choose to
(broken) program might depend on the old method. So ignore these rules and then complain when internal inter-
far there is only one way out: change the SONAME. faces change or go away entirely. There is no justification
for these complaints but developers can save themselves
With Linux’s symbol versioning mechanism this is not a lot of nerves by restricting the set of interfaces which
necessary. ABIs can normally be kept stable for as long are exported from the DSO.
as wanted. The symbol versioning mechanism [4] is an
extension of Sun’s internal versioning mechanism. The Section 2.2 introduced the mechanisms available for this.
two main differences are: it is possible to have more than Now that symbol versioning has been introduced we can
one definition of a given symbol (the associated version extend this discussion a bit. Using symbol maps was in-
must differ) and the application or DSO linked with the troduced as one of the possibilities to restrict the exported
versioned DSO contains not only a list of the required symbols. Ideally symbol maps should be used all the
version, but also records for each symbol which sym- time, in addition to the other method chosen. The rea-
bol version was used and from which DSO the definition son is that this allows associating version names with the
came. At runtime this information can then be used to interfaces which in turn later allow incompatible changes
pick the right version from all the different versions of to be made without breaking the ABI. The example map
the same interface. The only requirement is that the API file in section 2.2.5 does not define a version name, it is
(headers, DSO used for linking, and documentation) is an anonymous version map. The version defining a ver-
consistent. Every versioned DSO has at most one version sion name would look like this:
of every API which can be used at link-time. An API
(not ABI) can also vanish completely: this is a way to
deprecate APIs without affecting binary compatibility.
VERS_1.0 {
The only real problem of this approach is that if more global: index;
than one version of the same API is used in the same ap- local: *;
plication. This can only happen if the uses are in different };
objects, DSOs or the application itself. From inside one
object only one version can be accessed. In the last 8+
years this hasn’t been found to be a problem in the GNU
C library development. If it becomes a problem it can po- In this example VERS 1.0 is an arbitrarily chosen version
tentially be worked around by changing the implementa- name. As far as the static and dynamic linker are con-
tion of the multi-versioned interface to make it aware of cerned version names are simply strings. But for mainte-
it. Since both versions of the interface are implemented nance purposes it is advised that the names are chosen to
in the same DSO the versions can coordinate. In general, include the package name and a version number. For the
most of the implementation of the different versions is GNU C library project, for instance, the chosen names
shared and the actual versioned interfaces are normally are GLIBC 2.0, GLIBC 2.1, etc.
wrappers around a general implementation (see below).
12 Apparently some of the BSD variants “borrowed” the symbol ver-

sioning design. They never told me though.

36 Version 4.1.2 How To Write Shared Libraries

3.5 Handling Compatible Changes (GNU) symbol defined using @@ is the default definition. There
must be at most one. It is the version of the symbol used
in all linker runs involving the DSO. No symbol defined
The two basic compatible changes, extending functional- using @ are ever considered by the linker. These are the
ity of an existing interface and introducing a new inter- compatibility symbols which are considered only by the
face, can be handled similarly but not exactly the same dynamic linker.
way. And we need slightly different code to handle the
Linux/Hurd and the Solaris way. To exemplify the changes In this example we define both versions of the symbol to
we extend the example in section 2.2. The index func- use the same code. We could have just as well kept the
tion as defined cannot handle negative parameters. A ver- old definition of the function and added the new defini-
sion with this deficiency fixed can handle everything the tion. This would have increased the code size but would
old implementation can handle but not vice versa. There- provide exactly the same interface. Code which calls the
fore applications using the new interface should be pre- old version, index@VERS 1.0, would have produced un-
vented from running if only the old DSO is available. As specified behavior with the old DSO and now it would re-
a second change assume that a new function indexp1 is turn the same as a call to index@@VERS 2.0. But since
defined. The code would now look like this when using such a call is invalid anyway nobody can expect that the
Linux/Hurd: ABI does not change in this regard.

Since this code introduced a new version name the map

file has to change, too.
static int last;

static int next (void) {

return ++last;
VERS_1.0 {
}
global: index;
local: *;
int index1__ (int scale) {
};
return next () << (scale>0 ? scale : 0);
}
VERS_2.0 {
extern int index2__ (int)
global: index; indexp1;
__attribute ((alias ("index1__")));
} VERS_1.0;
asm(".symver index1__,index@VERS_1.0");
asm(".symver index2__,index@@VERS_2.0");

int indexp1 (int scale) {

The important points to note here are that index is men-
return index2__ (scale) + 1;
}
tioned in more than one version, indexp1 only appears
in VERS 2.0, the local: definitions only appear in the
VERS 1.0 definition, and the definition of VERS 2.0 refers
to VERS 1.0. The first point should be obvious: we want
Several things need explaining here. First, we do not ex- two versions of index to be available, this is what the
plicitly define a function index anymore. Instead index1 source code says. The second point is also easy to under-
is defined (note the trailing underscore characters; lead- stand: indexp1 is a new function and was not available
ing underscores are reserved for the implementation). This when version 1 of the DSO was released. It is not nec-
function is defined using the new semantic. The extern essary to mark the definition of indexp1 in the sources
declaration following the function definition is in fact a with a version name. Since there is only one definition
definition of an alias for the name index1 . This is the linker is able to figure this out by itself.
gcc’s syntax for expressing this. There are other ways
to express this but this one uses only C constructs which The omission of the catch-all local: * case might be
are visible to the compiler. The reason for having this a bit surprising. There is no local: case at all in the
alias definition can be found in the following two lines. VERS 2.0 definition. What about internal symbols in-
These introduce the “magic” to define a versioned sym- troduced in version 2 of the DSO? To understand this it
bol with more than one definition. The assembler pseudo- must be noted that all symbols matched in the local:
op .symver is used to define an alias of a symbol which part of the version definition do not actually get a version
consists of the official name, @ or @@, and a version name. name assigned. They get a special internal version name
The alias name will be the name used to access the sym- representing all local symbols assigned. So, the local:
bol. It has to be the same name used in the original code, part could appear anywhere, the result would be the same.
index in this case. The version name must correspond Duplicating local: * could possibly confuse the linker
to the name used in the map file (see the example in the since now there are two catch-all cases. It is no problem
previous section). to explicitly mention newly introduced local symbols in
the local: cases of new versions, but it is normally not
What remains to be explained is the use of @ and @@. The necessary since there always should be a catch-all case.

Ulrich Drepper Version 4.1.2 37

The fourth point, the VERS 1.0 version being referred removed as well, programs linked against the old DSO
to in the VERS 2.0 definition, is not really important in and referencing index in version VERS 1.0 would stop
symbol versioning. It marks the predecessor relationship working. Just like symbols, version names must never be
of the two versions and it is done to maintain the similar- removed.
ities with Solaris’ internal versioning. It does not cause
any problem it might in fact be useful to a human reader The code in this section has one little problem the code
so predecessors should always be mentioned. for the GNU versioning model in the previous section
does not have: the implementation of indexp1 refer-
3.6 Handling Compatible Changes (Solaris) ences the public symbol index and therefore calls it with
a jump to the PLT (which is slower and possibly allows
interposition). The solution for this is left as an exercise
The code changes to the code of the last section to handle to the user (see section 2.2.7).
Solaris’ internal versioning simplify sources and the map
file. Since there can only be one definition of a symbol The Solaris runtime linker uses the predecessor imple-
(and since a symbol cannot be removed there is exactly mentation to determine when it finds an interface not avail-
one definition) we do not need any alias and we do not able with the version found at link-time. If an applica-
have to mention index twice in the map file. The source tion was linked with the old DSO constructed from the
code would look like this: code above it would reference index@VERS 1.0. If the
new DSO is found at runtime the version found would
be index@VERS 2.0. In case such a mismatch is found
static int last; the dynamic linker looks into the list of symbols and tries
all predecessors in turn until all are checked or a match
static int next (void) { is found. In our example the predecessor of VERS 2.0
return ++last; is VERS 1.0 and therefore the second comparison will
} succeed.

int index (int scale) { 3.7 Incompatible Changes

return next () << (scale>0 ? scale : 0);
}
Incompatible changes can only be handled with the sym-
int indexp1 (int scale) { bol versioning mechanism present on Linux and GNU
return index (scale) + 1; Hurd. For Solaris one has to fall back to the filename
}
versioning method.

Having just covered the code for the compatible change,

Note that this only works because the previously defined the differences are not big. For illustration we pick up the
semantics of the index function is preserved in the new example code once again. This time, instead of making a
implementation. If this would not be the case this change compatible change to the semantics of index we change
would not qualify as compatible and the whole discussion the interface.
would be moot. The equally simplified map file looks like
this:
static int last;

static int next (void) {

VERS_1.0 {
return ++last;
local: *;
}
};
VERS_2.0 {
int index1__ (int scale) {
global: index; indexp1;
return next () << scale;
} VERS_1.0;
}
asm(".symver index1__,index@VERS_1.0");

int index2__ (int scale, int *result) {

The change consists of removing the index entry from if (result < 0
version VERS 1.0 and adding it to VERS 2.0. This leaves || result >= 8 * sizeof (int))
no exported symbol in version VERS 1.0 which is OK. return -1;
It would be wrong to remove VERS 1.0 altogether after *result = index1__ (scale);
moving the local: * case to VERS 2.0. Even if the return 0;
move would be done the VERS 1.0 definition must be }
kept around since this version is named as the predeces- asm(".symver index2__,index@@VERS_2.0");
sor of VERS 2.0. If the predecessor reference would be

38 Version 4.1.2 How To Write Shared Libraries

The interface of index in version VERS 2.0 as imple- one file, while the files needed at runtime are in another.
mented in index2 (note: this is the default version as
can be seen by the two @ in the version definition) is quite 3.8 Using Versioned DSOs
different and it can easily be seen that programs which
previously return some more or less sensible value now
can crash because *result is written to. This parameter All methods which depend on symbol versioning have
would contain garbage if the function would be used with one requirement in common: it is absolutely necessary
an old prototype. The index1 definition is the same as for the users of the DSO to always link with it. This might
the previous index implementation. We once again have sound like a strange requirement but it actually is not
to define the real function with an alias since the index since it is not necessary to provide definitions for all ref-
names get introduced by the .symver pseudo-ops. erences when creating DSOs. I.e., the linker is perfectly
happy if a symbol is completely undefined in a DSO. It
It is characteristic for incompatible changes that the im- is only the dynamic linker which would complain: if no
plementations of the different versions require separate object is in scope which defines the undefined symbol the
code. But as in this case one function can often be imple- lookup fails.
mented in terms of the other (the new code using the old
code as it is the case here, or vice versa). This method is sometimes, rarely, a useful method. If a
DSO is behaving differently depending on the context it
The map file for this example looks very much like the is loaded in and if the context has to provide some call-
one for the compatible change: backs which make the DSO complete, using undefined
symbols is OK. But this does not extend to definitions
from DSOs which use symbol versioning.

VERS_1.0 { The problem is that unless the DSO containing the defi-
global: index; nitions is used at link time, the linker cannot add a ver-
local: *;
sion name to the undefined reference. Following the rules
};
for symbol versioning [4] this means the earliest version
VERS_2.0 { available at runtime is used which usually is not the in-
global: index; tended version. Going back to the example in section 3.7,
} VERS_1.0; assume the program using the DSO would be compiled
expecting the new interface index2 . Linking happens
without the DSO, which contains the definition. The ref-
erence will be for index and not index@@VERS 2.0. At
We have two definitions of index and therefore the name runtime the dynamic linker will find an unversioned ref-
must be mentioned by the appropriate sections for the two erence and versioned definitions. It will then select the
versions. oldest definition which happens to be index1 . The re-
sult can be catastrophic.
It might also be worthwhile pointing out once again that
the call to index1 in index2 does not use the PLT It is therefore highly recommended to never depend on
and is instead a direct, usually PC-relative, jump. undefined symbols. The linker can help to ensure this if
-Wl,-z,defs is added to compiler command line. If it
With a simple function definition, like the one in this ex- is really necessary to use undefined symbols the newly
ample, it is no problem at all if different parts of the pro- built DSO should be examined to make sure that all ref-
gram call different versions of the index interface. The erences to symbols in versioned DSOs are appropriately
only requirement is that the interface the caller saw at marked.
compile time, also is the interface the linker finds when
handling the relocatable object file. Since the relocatable 3.9 Inter-Object File Relations
object file does not contain the versioning information it
is not possible to keep object files around and hope the
right interface is picked by the linker. Symbol versioning Part of the ABI is also the relationship between the var-
only works for DSOs and executables. If it is necessary ious participating executables and DSOs which are cre-
to reuse relocatable object files later, it is necessary to ated by undefined references. It must be ensured that
recreate the link environment the linker would have seen the dynamic linker can locate exactly the right DSOs at
when the code was compiled. The header files (for C, program start time. The static linker uses the SONAME
and whatever other interface specification exists for other of a DSO in the record to specify the interdependencies
languages) and the DSOs which are used for linking to- between two objects. The information is stored in the
gether form the API. It is not possible to separate the two DT NEEDED entries in the dynamic section of the object
steps, compiling and linking. For this reason packaging with the undefined references. Usually this is only a file
systems, which distinguish between runtime and devel- name, without a complete path. It is then the task of the
opment packages, put the headers and linkable DSOs in dynamic linker to find the correct file at startup time.

Ulrich Drepper Version 4.1.2 39

This can be a problem, though. By default the dynamic colons or a colon at the beginning or end of the string.
linker only looks into a few directories to find DSOs (on That means the run path value ”:/home::/usr:” searches
Linux, in exactly two directories, /lib and /usr/lib). the CWD, home, the CWD again, usr, and finally the
More directories can be added by naming them in the CWD again.13 It is very easy to add such an empty path.
/etc/ld.so.conf file which is always consulted by Makefiles often contain something like this:
the dynamic linker. Before starting the application, the
user can add LD LIBRARY PATH to the environment of
the process. The value is another list of directories which
are looked at. RPATH = $(GLOBAL\_RPATH):$(LOCAL\_RPATH)
LDFLAGS += -Wl,-rpath,$(RPATH)
But all this means that the selection of the directories is
under the control of the administrator. The program’s au-
thor can influence these setting only indirectly by doc-
If either GLOBAL RPATH or LOCAL RPATH is empty the
umenting the needs. But there is also way for the pro-
dynamic linker will be forced to look in the CWD. When
grammer to decide the path directly. This is sometimes
constructing strings one must therefore always be careful
important. The system’s setting might not be usable by
about empty strings.
all programs at the same time.
The second issue run paths have is that either that paths
For each object, DSO as well as executable, the author
like /usr/lib/someapp are not “relocatable”. I.e., the
can define a “run path”. The dynamic linker will use the
package cannot be installed in another place by the user
value of the path string when searching for dependencies
without playing tricks like creating symbolic links from
of the object the run path is defined in. Run paths comes
/usr/lib/someapp to the real directory. The use of rel-
is two variants, of which one is deprecated. The runpaths
ative paths is possible, but highly discouraged. It might
are accessible through entries in the dynamic section as
be OK in application which always control their CWD,
field with the tags DT RPATH and DT RUNPATH. The dif-
but in DSOs which are used in more than one application
ference between the two value is when during the search
using relative paths means calling for trouble since the
for dependencies they are used. The DT RPATH value is
application can change the CWD.
used first, before any other path, specifically before the
path defined in the LD LIBRARY PATH environment vari-
A solution out of the dilemma is an extension syntax for
able. This is problematic since it does not allow the user
all search paths (run paths, but also LD LIBRARY PATH).
to overwrite the value. Therefore DT RPATH is depre-
If one uses the string $ORIGIN this will represent the ab-
cated. The introduction of the new variant, DT RUNPATH,
solute path of the directory the file containing this run
corrects this oversight by requiring the value is used after
path is in. One common case for using this “dynamic
the path in LD LIBRARY PATH.
string token” (DST) is a program (usually installed in a
bin/ directory) which comes with one or more DSOs it
If both a DT RPATH and a DT RUNPATH entry are avail-
needs, which are installed in the corresponding lib/ di-
able, the former is ignored. To add a string to the run
rectory. For instance the paths could be /bin and /lib
path one must use the -rpath or -R for the linker. I.e.,
or /usr/bin and /usr/lib. In such a case the run
on the gcc command line one must use something like
path of the application could contain $ORIGIN/../lib
which will expand in the examples case just mentioned
gcc -Wl,-rpath,/some/dir:/dir2 file.o
to /bin/../lib and /usr/bin/../lib respectively.
The effective path will therefore point to the appropriate
This will add the two named directories to the run path
lib/ directory.
in the order in which say appear on the command line. If
more than one -rpath/-R option is given the parameters
$ORIGIN is not the only DST available. The GNU libc
will be concatenated with a separating colon. The order
dynamic currently recognizes two more. The first is $LIB
is once again the same as on the linker command line.
which is useful on platforms which can run in 32- and 64-
For compatibility reasons with older version of the linker
bit mode (maybe more in future). On such platforms the
DT RPATH entries are created by default. The linker op-
replacement value in 32-bit binaries is lib and for 64-
tion --enable-new-dtags must be used to also add
bit binaries it is lib64. This is what the system ABIs
DT RUNPATH entry. This will cause both, DT RPATH and
specify as the directories for executable code. If the plat-
DT RUNPATH entries, to be created.
form does not know more than one mode the replacement
value is lib. This DST makes it therefore easy to write
There are a number of pitfalls one has to be aware of
Makefiles creating the binaries since no knowledge about
when using run paths. The first is that an empty path
the differentiation is needed.
represents the current working directory (CWD) of the
process at runtime. One can construct such an empty path
The last DST is $PLATFORM. It is replaced by the dy-
by explicitly adding such a parameter (-Wl,-R,"") but
13 The dynamic linker is of course free to avoid the triple search since
also, and this is the dangerous part, by two consecutive
after the first one it knows the result.

40 Version 4.1.2 How To Write Shared Libraries

namic linker with an architecture-specific string repre-
senting the name of the platform (as the name suggests).
For instance, it will be replaced with i386 or i686 on
IA-32 machines, depending on the CPU(s) used. This al-
lows using better optimized versions of a DSO if the CPU
supports it. On system with the GNU libc this is gener-
ally not needed since the dynamic linker by itself and by
default looks for such subdirectories. The actual rules are
quite complicated and not further discussed here. What
is important to remember is that the $PLATFORM DST is
not really useful; it is mainly available for compatibility
with Solaris’s dynamic linker.

Another aspect of DT NEEDED entries worth looking at

is whether they are necessary in the first place. Espe-
cially when the above guideline of using the -z defs
linker option is followed, many projects name all kinds of
DSOs on the linker command line, whether they are re-
ally needed or not in the specific case. This problem is in-
tensified with useless wrappers like pkg-config which
just add tons of dependencies to the link time. This might
have been OK in the days of static linking, but the de-
fault for DSOs appearing on the linker command line
is, that they are always added to the result’s dependency
list, whether they are actually used or not. To determine
whether an executable or DSO has such unnecessary de-
pendencies the ldd script can be used:

$ ldd -u -r \
/usr/lib/libgtk-x11-2.0.so.0.600.0
Unused direct dependencies:

/usr/lib/libpangox-1.0.so.0
/lib/libdl.so.2

These references can be manually eliminated by avoiding

to name the DSO on the linker command line. Alterna-
tively the --as-needed linker option can be used. If
this option is used, all DSOs named on the command line
after the option are only added to the dependency list if
they are really needed. This mode can be disabled again
with the --no-as-needed option. This way the lazy
programmer can still name all the DSOs which might be
needed at all times. The linker does all the hard work.

This process should not be executed mindlessly, though.

Not having the DSO on the direct dependency list any-
more means the symbol lookup path is altered. If the
DSO in question contains a definition of a symbol which
also appears in a second DSO and now that second DSO
is earlier in the lookup path, the program’s semantics
might be altered. This is a rare occurrence, though.

Ulrich Drepper Version 4.1.2 41

A Counting Relocations

The following script computes the number of normal and relative relocations as well as the number of PLT entries
present in a binary. If an appropriate readelf implementation is used it can also be used to look at all files in an
archive. If prelink [7] is available and used, the script also tries to provide information about how often the DSO is
used. This gives the user some idea how much “damage” an ill-written DSO causes.

#! /usr/bin/perl
eval "exec /usr/bin/perl -S $0 $*"
if 0;
# Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005 Ulrich Drepper
# Written by Ulrich Drepper <[email protected]>, 2000.

# This program is free software; you can redistribute it and/or modify

# it under the terms of the GNU General Public License version 2 as
# published by the Free Software Foundation.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program; if not, write to the Free Software Foundation,
# Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. */

for ($cnt = 0; $cnt <= $#ARGV; ++$cnt) {

$relent = 0;
$relsz = 0;
$relcount = 0;
$pltrelsz = 0;
$extplt = 0;
$users = 0;

open (READLINK, "readlink -f $ARGV[$cnt] |") || die "cannot open readlink";

while (<READLINK>) {
chop;
$fullpath = $_;
}
close (READLINK);

open (READELF, "eu-readelf -d $ARGV[$cnt] |") || die "cannot open $ARGV[$cnt]";

while (<READELF>) {
chop;
if (/.* \(RELENT\) *([0-9]*).*/) {
$relent = $1 + 0;
} elsif (/.* \(RELSZ\) *([0-9]*).*/) {
$relsz = $1 + 0;
} elsif (/.* \(RELCOUNT\) *([0-9]*).*/) {
$relcount = $1 + 0;
} elsif (/.* \(PLTRELSZ\) *([0-9]*).*/) {
$pltrelsz = $1 + 0;
}
}
close (READELF);

open (READELF, "eu-readelf -r $ARGV[$cnt] |") || die "cannot open $ARGV[$cnt]";

while (<READELF>) {
chop;
if (/.*JU?MP_SLOT *0+ .*/) {
++$extplt;
}

42 Version 4.1.2 How To Write Shared Libraries

 }
close (READELF);

if (open (PRELINK, "/usr/sbin/prelink -p 2>/dev/null | fgrep \"$fullpath\" |")) {

while (<PRELINK>) {
++$users;
}
close(PRELINK);
} else {
$users = -1;
}

printf("%s: %d relocations, %d relative (%d%%), %d PLT entries, %d for local syms (%d%%)",
$ARGV[$cnt], $relent == 0 ? 0 : $relsz / $relent, $relcount,
$relent == 0 ? 0 : ($relcount * 100) / ($relsz / $relent),
$relent == 0 ? 0 : $pltrelsz / $relent,
$relent == 0 ? 0 : $pltrelsz / $relent - $extplt,
$relent == 0 ? 0 : ((($pltrelsz / $relent - $extplt) * 100)
/ ($pltrelsz / $relent)));
if ($users >= 0) {
printf(", %d users", $users);
}
printf("\n");
}

Ulrich Drepper Version 4.1.2 43

B Automatic Handler of Arrays of String Pointers

The method to handle arrays of string pointers presented in section 2.4.3 show the principle method to construct data
structures which do not require relocations. But the construction is awkward and error-prone. Duplicating the strings
in multiple places in the sources always has the problem of keeping them in sync.

Bruno Haible suggested something like the following to automatically generate the tables. The programmer only has
to add the strings, appropriately marked, to a data file which is used in the compilation. The framework in the actual
sources looks like this:

#include <stddef.h>

#define MSGSTRFIELD(line) MSGSTRFIELD1(line)

#define MSGSTRFIELD1(line) str##line
static const union msgstr_t {
struct {
#define _S(n, s) char MSGSTRFIELD(__LINE__)[sizeof(s)];
#include "stringtab.h"
#undef _S
};
char str[0];
} msgstr = { {
#define _S(n, s) s,
#include "stringtab.h"
#undef _S
} };
static const unsigned int msgidx[] = {
#define _S(n, s) [n] = offsetof(union msgstr_t, MSGSTRFIELD(__LINE__)),
#include "stringtab.h"
#undef _S
};
const char *errstr (int nr) {
return msgstr.str + msgidx[nr];
}

The string data has to be provided in the file stringtab.h. For the example from section 2.4.3 the data would look
like this:

_S(ERR1, "message for err1")

_S(ERR3, "message for err3")
_S(ERR2, "message for err2")

The macro S takes two parameters: the first is the index used to locate the string and the second is the string itself.
The order in which the strings are provided is not important. The value of the first parameter is used to place the offset
in the correct slot of the array. It is worthwhile running these sources through the preprocessor to see the results. This
way of handling string arrays has the clear advantage that strings have to be specified only in one place and that the
order they are specified in is not important. Both these issues can otherwise easily lead to very hard to find bugs.

The array msgidx in this cases uses the type unsigned int which is in most cases 32 bits wide. This is usually far
too much to address all bytes in the string collectionin msgstr. So, if size is an issue the type used could be uint16 t
or even uint8 t.

Note that both arrays are marked with const and therefore are not only stored in the read-only data segment and there-
fore shared between processes, preserving, precious data memory, making the data read-only also prevents possible
security problems resulting from overwriting values which can trick the program into doing something harmful.

44 Version 4.1.2 How To Write Shared Libraries

C Index
copy relocation, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
--as-needed, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 cost of relocations, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
--hash-style, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 CPU cache, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
--no-as-needed, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
--shared, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 deep binding, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
--version-script, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 default symbol, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
-fno-common, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 default visibility, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18ff.
-fpic, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, 15 destructor, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
-fvisibility, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 DF BIND NOW, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
-fvisibility-inlines-hidden, . . . . . . . . . . . . . . . . . 21 DF DYNAMIC, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
-rpath, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 DF SYMBOLIC, . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10, 20, 25
-z defs, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39, 41 DF TEXTREL, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
-z now, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5, 32 DL CALL FCT, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
-z relro, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 dlopen(), . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10, 33
.symver, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37f. dlsym(), . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
@GOT, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, 17, 23 documented interface, . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
@GOTOFF, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18, 23, 30 DSO, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
@PLT, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, 17 DST, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
$LIB, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 DT FLAGS, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5, 25
$ORIGIN, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 DT GNU HASH, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
$PLATFORM, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 DT NEEDED, . . . . . . . . . . . . . . . . . . . . . . . . . . . 10, 35, 39, 41
DT RPATH, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
ABI, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34, 36 DT RUNPATH, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
versioning, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 DT SONAME, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
absolute address, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 DT SYMBOLIC, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20, 25
access permission, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 DT TEXTREL, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Ada names, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 dynamic linker, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2, 4
adding symbols, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 dynamic shared object, . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
address space, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 dynamic string token, . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
aliases, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24, 37 dynamic symbol table, . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
alternative entry point, . . . . . . . . . . . . . . . . . . . . . . . . . . 24
anonymous version map, . . . . . . . . . . . . . . . . . . . . . . . . 36 ELF, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
a.out, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1f. entry point, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
API, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34, 36 exception handling, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
auxiliary vector, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 executable linkage format, . . . . . . . . . . . . . . . . . . . . . . . . 2
execve(), . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
binary format, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 export map, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Bloom filter, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
BSS section, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Fedora Core, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
file-local scope, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
C++ filename versioning, . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
classes, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 fini, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
names, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 function array, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
typeinfo, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 function attribute, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
virtual function table, . . . . . . . . . . . . . . . . . . . . . . . 29 function semantics, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
cache prefetching, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
chain array, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 gABI, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
chain length, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 global offset table, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
class visibility, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 GNU, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
code generation, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 versioning, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37f.
COFF, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 GOT, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, 32
common variable, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 attacking, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
compatibility, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 gprof, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
compatibility symbol, . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
compatible changes, . . . . . . . . . . . . . . . . . . . . . . . . . 35, 37 hash
computed goto, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 chain, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6, 9
constant data, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27, 32 collisions, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
constructor, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 function, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

Ulrich Drepper Version 4.1.2 45

hidden visibility, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18ff. relocation, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Hurd, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36f. counting, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
RTLD DEEPBIND, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
IA-32, . . . . . . . . . . . . . . . . . . . . . . . 11f., 15, 17f., 30ff., 34 RTLD GLOBAL, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
IA-64, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31f. run path, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
implementation-defined behavior, . . . . . . . . . . . . . . . . 34
incompatible changes, . . . . . . . . . . . . . . . . . . . . . . . 36, 38 scope, lookup, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
init, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 security, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12, 32
initialized variable, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 Security Enhanced Linux, . . . . . . . . . . . . . . . . . . . . . . . 33
interposition, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6, 12 segment, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
SELinux, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
lazy relocation, . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2, 5, 32 shared library, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
LD BIND NOW, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5, 35 shared object name, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
ldd, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 SHF MERGE, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
LD DEBUG, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8, 14 SHF STRINGS, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
LD LIBRARY PATH, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 Solaris, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22, 35–38
LD PRELOAD, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 SONAME, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33, 35f., 39
LD PROFILE, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25, 33 sprof, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
libtool, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2, 23 stable interface, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Linux,. . . . . . . . . . . . . . . . . . . . . . . .1f., 7, 12, 33, 36f., 39 startup costs, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
load address, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1ff. static linking, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
lookup process, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 STV HIDDEN, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
lookup scope, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5, 10 string definitions, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
STV DEFAULT, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
member function, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 successful lookup, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
memory page, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 switch, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
memory protection, . . . . . . . . . . . . . . . . . . . . . . . . . 3f., 32 symbol
mmap(), . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
alias, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
mprotect(), . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
export, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17, 36
reference, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
name mangling, . . . . . . . . . . . . . . . . . . . . . . . . . 7, 22, 25f.
relocation, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5, 30
non-lazy relocation, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
table, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9, 13
object file, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 versioning, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35f.
OpenOffice.org, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8, 10 visibility, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
overhead, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 weak, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
symbol table, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
PC-relative, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18, 23, 31
PIC register, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11, 31 text relocation, . . . . . . . . . . . . . . . . . . . . . . . . 5, 15, 31, 33
PIE, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2, 33 text segment, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
PLT, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
attacking, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 undefined behavior, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
pointer array, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 undefined symbol, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
position independent executable, . . . . . . . . . . . . . . . 2, 33 uninitialized variable, . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
pragma visibility, . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19, 21 unsuccessful lookup, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
pre-linking, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4, 9
versioning, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
prefetching, cache, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
procedure linkage table, . . . . . . . . . . . . . . . . . . . . . . . . . 11 weak symbol definition, . . . . . . . . . . . . . . . . . . . . . . . . . . 6
profiling, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 wrapper function, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
program header, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 write access, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
psABI, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2, 12
PT GNU RELRO, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 x86-64, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31f.
PT INTERP, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
PT LOAD, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3f.

read-only memory, . . . . . . . . . . . . . . . . . . . . . . . . . . . 5, 27
relative address, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
relative relocation, . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5, 30
relinfo, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
relocatable binary, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
relocatable object file, . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

46 Version 4.1.2 How To Write Shared Libraries

D References
[1] System V Application Binary Interface, http://www.caldera.com/developers/gabi/, 2001.
[2] Ulrich Drepper, Red Hat, Inc., ELF Handling For Thread-Local Storage, http://people.redhat.com/drepper/tls.pdf,
2003.

[3] Ulrich Drepper, Red Hat, Inc., Good Practices in Library Design, Implementation, and Maintenance,
http://people.redhat.com/drepper/goodpractices.pdf, 2002.
[4] Ulrich Drepper, Red Hat, Inc., ELF Symbol Versioning, http://people.redhat.com/drepper/symbol-versioning, 1999.
[5] Sun Microsystems, Linker and Library Guide, http://docs.sun.com/db/doc/816-1386, 2002.

[6] TIS Committee, Executable and Linking Format (ELF) Specification, Version 1.2,
http://x86.ddj.com/ftp/manuals/tools/elf.pdf, 1995.
[7] Jakub Jelinek, Red Hat, Inc., Prelink, http://people.redhat.com/jakub/prelink.pdf, 2003.
[8] Security Enhanced Linux http://www.nsa.gov/selinux/.

E Revision History

2002-11-2 First public draft.

2002-11-8 Fixed a couple of typos.
Document one more possibility for handling arrays of string pointers.
Describe PLT on IA-32 in more details.

2002-11-14 Document implications of using C++ virtual functions.

2003-1-20 Incorporate several suggestions by Bruno Haible.
Describe visibility attribute and aliasing in C++ code.
2003-2-9 Some more cleanups. Version 1.0 release.

2003-2-27 Minor language cleanup. Describe using export maps with C++. Version 1.1.
2003-3-18 Some more linguistic changes. Version 1.2.
2003-4-4 Document how to write constructor/destructors. Version 1.3.

2003-12-8 Describe how to use run paths. Version 1.4.

2003-12-9 Add section about avoided PIC reload. Version 1.5.
2004-2-4 Fix some typos. Explain optimizations gcc does without -fpic. Explain -z relro. Version 1.7.
2004-2-8 Introduce the lookup scope in more details. Version 1.9.

2004-8-4 Warn about aliases of static objects. Significant change to section 2.2 to introduce new features of gcc 4.0.
Version 1.99.
2004-8-27 Update code in appendices A and B. Version 2.0.
2004-9-23 Document RTLD DEEPBIND and sprof. Version 2.2.

2005-1-22 Brushed up language in many places. Add Index. Version 3.0.

2006-8-20 Document GNU-style hash table. Version 4.0.

Ulrich Drepper Version 4.1.2 47